Composition optimization device, composition optimization method, and composition optimization program

Abstract

The purpose of the present invention is to efficiently search for a material composition corresponding to a target phase fraction. A composition optimization device includes: a prediction unit that predicts a phase fraction at each temperature of a material having a predetermined material composition within a predetermined temperature interval by inputting the predetermined material composition into a trained model trained using training data in which the phase fraction of the material having the predetermined material composition is greater than the phase fraction of the material at each temperature within the predetermined temperature interval; the material composition of the training object material is associated with the phase fractions of the training object material at each temperature within a predetermined temperature interval; and an updating unit that updates a predetermined material composition input in the trained model by back-propagating an error calculated on the basis of the target phase fraction and the predicted phase fraction, the process of predicting the phase fraction by the prediction unit on the basis of the updated material composition and the process of updating the predetermined material composition by the update unit by back-propagating the calculated error are repeated until the calculated error satisfies a predetermined condition.

Description

Technical Field

[0001] The invention relates to a composition optimization device, a composition optimization method and a composition optimization program. Background Art

[0002] In the design of materials such as alloys, it is important to calculate the phase fraction in a thermodynamic equilibrium state within a predetermined temperature range. As an example, a model for predicting the phase fraction from the material composition is known. By using this model, it is possible to predict how the phase fraction changes when the material composition is changed, for example.

[0003] Patent Document 1: International Publication No. 2020 / 090617 Summary of the invention

[0004] <Problems to be Solved by the Invention>

[0005] On the other hand, in the case of the above-mentioned model, the number of model parameters is large, and the prediction of the phase fraction requires a certain amount of time. Therefore, if a large number of forward problems are solved while the material composition is completely changed in order to explore the material composition corresponding to the target phase fraction (i.e., the inverse problem must be solved), it is impossible to explore the appropriate material composition within a realistic time.

[0006] An object of the present invention is to efficiently search for a material composition corresponding to a target phase fraction.

[0007] <Methods used to solve the problem>

[0008] The composition optimization device of the first aspect of the present invention comprises:

[0009] a prediction unit that predicts the phase fraction of a material having the predetermined material composition at each temperature within a predetermined temperature range by inputting a predetermined material composition into a trained model trained using training data in which the material composition of a training target material and the phase fraction of the training target material at each temperature within the predetermined temperature range are associated; and

[0010] an updating unit that updates the predetermined material composition input into the trained model by back-propagating an error calculated based on the target phase fraction and the predicted phase fraction,

[0011] The process of predicting the phase fraction based on the updated material composition by the prediction unit and the process of updating the predetermined material composition by back-propagating the calculated error by the update unit are repeatedly performed until the calculated error satisfies a predetermined condition.

[0012] The second aspect of the present invention is a composition optimization device according to the first aspect, wherein:

[0013] The trained model is configured to predict the phase fraction at the (i+1)th temperature using the phase fractions up to the i-th temperature within the predetermined temperature range predicted by the trained model, where i is an integer greater than or equal to 1.

[0014] The third aspect of the present invention is a composition optimization device according to the second aspect, wherein:

[0015] The trained model adopts a structure capable of calculating data at predetermined time intervals, namely, time series data. The trained model predicts the phase fraction at predetermined temperature intervals based on the predetermined material composition of the material.

[0016] The fourth aspect of the present invention is a composition optimization device according to the third aspect, wherein:

[0017] The trained model is any one of RNN, bidirectional RNN, Seq2Seq, Seq2Seq with Attention mechanism, GRU, LSTM and Transformer.

[0018] The fifth aspect of the present invention is a composition optimization device according to the fourth aspect, wherein:

[0019] The trained model includes:

[0020] an encoder unit that outputs a feature amount by inputting a material composition of a predetermined material; and

[0021] A decoder unit predicts the phase fraction at the (i+1)th temperature by inputting the output feature quantity and the predicted phase fractions up to the i-th temperature.

[0022] The sixth aspect of the present invention is a composition optimization device according to any one of aspects 1 to 5, wherein:

[0023] The phase fraction is the phase fraction in a thermodynamic equilibrium state.

[0024] The seventh aspect of the present invention is a composition optimization device according to any one of aspects 1 to 6, wherein:

[0025] The invention further includes: a constraint unit that corrects the updated predetermined material composition in such a manner that the updated predetermined material composition updated by the updating unit satisfies the constraint related to the material composition.

[0026] The eighth aspect of the present invention is a composition optimization device according to the seventh aspect, wherein:

[0027] The constraint unit corrects the updated predetermined material composition so that the total value of the material composition is 100% and the ratio of each composition is 0% or more.

[0028] The ninth aspect of the present invention is a composition optimization device according to any one of the first to eighth aspects, wherein:

[0029] The system further includes a loss function calculation unit that calculates an error calculated based on the target phase fraction and the predicted phase fraction.

[0030] A tenth aspect of the present invention is a composition optimization device according to the ninth aspect, wherein:

[0031] The error calculated by the loss function calculation unit includes at least any one of the following:

[0032] a first addition result obtained by adding the predicted phase fraction at the temperature at which each phase is generated or disappeared determined based on the target phase fraction for all phases and an error between the target phase fraction;

[0033] a second addition result obtained by adding the error between the predicted phase fraction at each temperature within a predetermined temperature range and the target phase fraction at each temperature;

[0034] a third addition result obtained by adding the error between the predicted logarithmic value of the phase fraction at each temperature within a predetermined temperature range and the target logarithmic value of the phase fraction at each temperature;

[0035] a fourth addition result obtained by adding a difference value of the phase fraction between adjacent temperatures in the predicted phase fraction at each temperature within a predetermined temperature range and an error between the difference value of the phase fraction between adjacent temperatures in the predicted phase fraction at each temperature; and

[0036] The fifth addition result is obtained by adding the error between the predicted ratio of the phase fractions at each temperature within a predetermined temperature range and the target ratio of the phase fractions at each temperature.

[0037] The 11th aspect of the present invention is a composition optimization device according to the 10th aspect, wherein:

[0038] The loss function calculation unit performs weighted addition on the first to fifth addition results.

[0039] The twelfth aspect of the present invention is a composition optimization device according to the tenth aspect, wherein:

[0040] When calculating the logarithmic value of the phase fraction, the loss function calculation unit makes the logarithmic value of the phase fraction a non-negative value by adding a value corresponding to the first decimal place of the phase fraction.

[0041] The 13th aspect of the present invention is a composition optimization device according to any one of the 1st to 12th aspects, wherein:

[0042] The predetermined material composition is a material composition selected from material compositions specified by a standard and having a phase fraction similar to a target phase fraction.

[0043] A 14th aspect of the present invention is a composition optimization method, wherein the following steps are performed by a computer of a composition optimization device:

[0044] a prediction step of predicting the phase fraction of a material having the predetermined material composition at each temperature within a predetermined temperature range by inputting a predetermined material composition into a trained model trained using training data in which the material composition of a training target material and the phase fraction of the training target material at each temperature within the predetermined temperature range are associated; and

[0045] an updating step of updating the predetermined material composition input into the trained model by back-propagating an error calculated based on the target phase fraction and the predicted phase fraction,

[0046] The process of predicting the phase fraction based on the updated material composition in the prediction step and the process of updating the predetermined material composition by back-propagating the calculated error in the update step are repeatedly performed until the calculated error satisfies a predetermined condition.

[0047] A fifteenth aspect of the present invention is a composition optimization program that causes a computer of a composition optimization device to execute the following steps:

[0048] a prediction step of predicting the phase fraction of a material having the predetermined material composition at each temperature within a predetermined temperature range by inputting a predetermined material composition into a trained model trained using training data in which the material composition of a training target material and the phase fraction of the training target material at each temperature within the predetermined temperature range are associated; and

[0049] an updating step of updating the predetermined material composition input into the trained model by back-propagating an error calculated based on the target phase fraction and the predicted phase fraction,

[0050] The process of predicting the phase fraction based on the updated material composition in the prediction step and the process of updating the predetermined material composition by back-propagating the calculated error in the update step are repeatedly performed until the calculated error satisfies a predetermined condition.

[0051] <Effects of the Invention>

[0052] According to the present invention, it is possible to efficiently search for a material composition corresponding to a target phase fraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a diagram showing an example of the system configuration of a material development support system and an example of the functional configuration of a training device, a prediction device, and a composition optimization device.

[0054] Figure 2 This is a diagram showing an example of the hardware configuration of a training device, a prediction device, and a composition optimization device.

[0055] Figure 3 This is a flowchart showing the flow of material development support processing.

[0056] Figure 4 It is a diagram showing a specific example of the structure of training data and input data and correct answer data.

[0057] Figure 5 This is a diagram showing an example of the functional structure of the training data generating unit.

[0058] Figure 6 This is a diagram showing an example of the functional structure of a training unit.

[0059] Figure 7 FIG. 1 is a first diagram showing an operation example of the training unit.

[0060] Figure 8 FIG. 2 is a second diagram showing an operation example of the training unit.

[0061] Fig. 9 This is a diagram showing an example of a method for calculating a loss performed by a loss function calculation unit.

[0062] Fig.10 It is a flowchart which shows the flow of a training process.

[0063] Fig.11 This is a diagram showing an example of the functional structure of a prediction unit.

[0064] Fig.12 It is a diagram showing an operation example of the prediction unit.

[0065] Fig.13 It is a flowchart which shows the flow of prediction processing.

[0066] Fig.14 It is a graph for explaining the prediction accuracy.

[0067] Fig.15 FIG1 is the first figure showing an operation example of the forward propagation processing in the composition optimization device.

[0068] Fig.16 FIG2 is a second diagram showing an operation example of the forward propagation processing in the composition optimization device.

[0069] Fig.17 FIG3 is a third diagram showing an operation example of the forward propagation processing in the composition optimization device.

[0070] Fig.18 It is a diagram showing an operation example of the back-propagation processing in the composition optimization device.

[0071] Fig.19 FIG. 4 is a fourth diagram showing an operation example of the forward propagation processing in the composition optimization device.

[0072] Fig. 20 It is a flowchart which shows the flow of composition optimization processing.

[0073] Fig.21 This is a flowchart showing the flow of material composition update processing.

[0074] Fig. 22 FIG1 is a first diagram showing an example of a processing result of the composition optimization processing.

[0075] Fig.23 FIG1 is another example of the functional structure of the composition optimization device.

[0076] Fig.24 FIG. 2 is another example of the functional structure of the composition optimization device.

[0077] Fig.25 This is a diagram showing an example of default material composition information.

[0078] Fig.26 This is a diagram showing an example of default material composition information and a processing result of composition optimization processing.

[0079] Fig. 27 FIG2 is a second diagram showing an example of a processing result of the composition optimization processing. DETAILED DESCRIPTION

[0080] Hereinafter, each embodiment will be described with reference to the drawings. It should be noted that in the specification and the drawings, components having substantially the same functional structure are given the same reference numerals to omit repeated descriptions.

[0081] [First embodiment]

[0082] <System structure of the material development support system, and functional structure of the training device, prediction device, and composition optimization device>

[0083] First, the system structure of the material development support system including the training device, prediction device, and composition optimization device of the first embodiment, and the functional structure of the training device, prediction device, and composition optimization device are described.

[0084] Predicting the phase fraction (phase fraction in thermodynamic equilibrium, hereinafter simply referred to as “phase fraction”) at each temperature within a predetermined temperature range based on the material composition information by using the trained prediction model (analytical problem), and

[0085] ·Explore the material composition information corresponding to the target phase fraction at each temperature within a predetermined temperature range by using the trained prediction model (analytical inverse problem)

[0086] system.

[0087] In the material development support system according to the present embodiment, the ratio of phase precipitation of each composition at each temperature is provided in the form of a curve graph, for example, represented in a predetermined temperature range. In this case, for materials such as aluminum alloys, the material strength can be judged based on the shape of the curve graph of the phase fraction. It should be noted that the phase fraction can also be provided as a numerical value of the ratio of phase precipitation of each composition at each temperature.

[0088] Figure 1 FIG. 1 is a diagram showing an example of a system configuration of a material development support system and an example of a functional configuration of a training device, a prediction device, and a composition optimization device. Figure 1 As shown, the material development support system 100 includes a training device 110 , a prediction device 120 and a composition optimization device 130 .

[0089] The training program is installed in the training device 110 , and when the program is executed, the training device 110 functions as the training data generating unit 111 and the training unit 112 .

[0090] The training data generating unit 111 generates training data for training the prediction model and stores the training data in the training data storing unit 113. In the present embodiment, as the training data for training the prediction model, in the training data storing unit 113, a plurality of material composition information is stored in association with the phase fraction at each temperature within a predetermined temperature range for each material composition information.

[0091] The training unit 112 reads the training data from the training data storage unit 113, and uses the read training data to train the prediction model. The training unit 112 trains the prediction model so that the output data when the material composition information stored in the training data is input into the prediction model is close to the phase fraction at each temperature in the predetermined temperature range stored in association with the training data. Thus, the training unit 112 generates a trained prediction model. In addition, the training unit 112 stores the generated trained prediction model in the prediction unit 122 of the prediction device 120 and the prediction unit 132 of the composition optimization device 130.

[0092] It should be noted that the training unit 112 uses the following architecture (generally used for natural language processing, etc.) that can calculate data at predetermined time intervals (ie, time series data) as a prediction model:

[0093] Seq2Seq with Attention, or

[0094] Transformer.

[0095] Therefore, according to this prediction model, by inputting the material composition information stored in the training data, it is possible to sequentially output output data corresponding to the phase fraction at each temperature within a predetermined temperature range.

[0096] Here, “outputting output data corresponding to the phase fraction at each temperature in sequence” means that, for example, the prediction model outputs output data corresponding to the phase fraction at the i+1th temperature using the correct answer data stored in the training data up to the i-th temperature, where i is an integer greater than 1.

[0097] As described above, the training unit 112 is configured to sequentially output output data corresponding to the phase fraction at each temperature (configured to process the output data corresponding to the phase fraction at each temperature as time series data). Thus, according to the training unit 112, training that reflects the output data corresponding to the phase fraction in adjacent temperature regions can be performed.

[0098] A prediction program is installed in the prediction device 120 , and by executing the program, the prediction device 120 functions as a material composition input unit 121 , a prediction unit 122 , and a display unit 123 .

[0099] When the material composition information of the prediction target material is input, the material composition input unit 121 receives the information and notifies the prediction unit 122 of the information.

[0100] The prediction unit 122 has a trained prediction model trained by the training unit 112, and sequentially predicts the phase fraction at each temperature in the predetermined temperature range by inputting the material composition information notified by the material composition input unit 121 into the trained prediction model. That is, the prediction unit 122 sequentially predicts the phase fraction in the predetermined temperature range at the predetermined temperature interval.

[0101] Here, "predicting the phase fraction sequentially at predetermined temperature intervals" means, for example, using the phase fraction predicted by a trained prediction model up to the i-th temperature (i is an integer greater than 1), and the trained prediction model predicts the phase fraction at the i+1-th temperature.

[0102] In this way, the prediction unit 122 is configured to sequentially predict the phase fraction at each temperature (that is, configured as a recursive network that processes the phase fraction at each temperature as time series data). Thus, according to the prediction unit 122, it is possible to predict the prediction data reflecting the phase fraction in the adjacent temperature range. As a result, compared with the case where a multi-layer neural network is applied (the case where the prediction result of the phase fraction in the adjacent temperature range is not reflected), the decrease in the prediction accuracy of the phase fraction can be suppressed. That is, according to the prediction unit 122, in the trained prediction model that predicts the phase fraction in a predetermined temperature range based on the material composition, the prediction accuracy can be improved.

[0103] The display unit 123 displays the phase fraction at each temperature within the predetermined temperature range predicted by the prediction unit 122. In the display unit 123, for example, each phase is distinguished by color to display the phase fraction at each temperature within the predetermined temperature range predicted by the prediction unit 122. It should be noted that the display of the phase fraction as a graph may also be displayed by changing the type of line such as a dotted line or a dashed line.

[0104] A composition optimization program is installed in the composition optimization device 130 . By executing the program, the composition optimization device 130 functions as a material composition input unit 131 , a prediction unit 132 , a loss function calculation unit 133 , and an update unit 134 .

[0105] When default material composition information is input, material composition input unit 131 receives the default material composition information and notifies prediction unit 132. When material composition information updated by update unit 134 is input, material composition input unit 131 receives the default material composition information and notifies prediction unit 132.

[0106] The prediction unit 132 has a trained prediction model trained by the training unit 112, and sequentially predicts the phase fraction at each temperature in the predetermined temperature range by inputting the material composition information notified by the material composition input unit 131 into the trained prediction model. That is, the prediction unit 132 sequentially predicts the phase fraction in the predetermined temperature range according to the predetermined temperature interval.

[0107] The loss function calculation unit 133 calculates a plurality of losses (errors) based on the phase fractions at each temperature within the predetermined temperature interval sequentially predicted by the prediction unit 132 and the target phase fractions at each temperature within the predetermined temperature interval set in advance. In addition, the loss function calculation unit 133 integrates the calculated plurality of losses to calculate the comprehensive loss, and performs back propagation on the calculated comprehensive loss while fixing the model parameters of the trained prediction model of the prediction unit 132.

[0108] The updating unit 134 updates the material composition information based on the back propagation result output from the trained prediction model of the prediction unit 132 by back propagating the comprehensive loss using the loss function calculation unit 133. In addition, the updating unit 134 inputs the updated material composition information to the material composition input unit 131.

[0109] As described above, when searching for a material composition corresponding to a target phase fraction, the composition optimization device 130 changes the material composition by back-propagating the comprehensive loss instead of changing the material composition comprehensively. Thus, the composition optimization device 130 can efficiently search for a material composition corresponding to a target phase fraction.

[0110] <Hardware structure of training device, prediction device and composition optimization device>

[0111] Next, the hardware structures of the training device 110, the prediction device 120, and the composition optimization device 130 are described. It should be noted that since the training device 110, the prediction device 120, and the composition optimization device 130 have the same hardware structure, Figure 2 The hardware structures of the training device 110, the prediction device 120 and the optimization device 130 are summarized and described.

[0112] Figure 2 FIG. 1 is a diagram showing an example of a hardware structure of a training device, a prediction device, and an optimization device. Figure 2 As shown, the training device 110, the prediction device 120 and the composition optimization device 130 include a processor 201, a memory 202, an auxiliary storage device 203, an I / F (interface) device 204, a communication device 205 and a drive device 206. It should be noted that the hardware of the training device 110, the prediction device 120 and the composition optimization device 130 are connected to each other via a bus 207.

[0113] The processor 201 includes various computing devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 201 reads various programs (eg, training programs, prediction programs, composition optimization programs, etc.) to the memory 202 and executes them.

[0114] The memory 202 includes a main storage device such as a ROM (Read Only Memory) and a RAM (Random Access Memory). The processor 201 and the memory 202 form a so-called computer, and the processor 201 executes various programs read out to the memory 202, and the computer realizes various functions.

[0115] The auxiliary storage device 203 stores various programs and various data used when the processor 201 executes the various programs. For example, the training data storage unit 113 is implemented in the auxiliary storage device 203.

[0116] The I / F device 204 is a connection device connected to an operation device 211 as an example of a user interface device and a display device 212. The communication device 205 is a communication device for communicating with an external device (not shown) via a network.

[0117] The drive device 206 is a device for placing a recording medium 213. The recording medium 213 includes a medium that records information optically, electrically, or magnetically, such as a CD-ROM, a floppy disk, a magneto-optical disk, etc. In addition, the recording medium 213 may include a semiconductor memory that records information electrically, such as a ROM, a flash memory, etc.

[0118] It should be noted that the various programs installed in the auxiliary storage device 203 are installed by, for example, placing the distributed recording medium 213 in the drive device 206 and reading the various programs recorded in the recording medium 213 through the drive device 206. Alternatively, the various programs installed in the auxiliary storage device 203 may be installed by downloading from a network via the communication device 205.

[0119] <Process of Material Development Assistant>

[0120] Next, the entire material development support process of the material development support system 100 will be described. Figure 3 It is a flowchart showing the process of the material development assistant.

[0121] In step S301, the training device 110 performs training processing to generate a trained prediction model.

[0122] In step S302, the prediction device 120 and the composition optimization device 130 store the trained prediction model.

[0123] In step S303 , the prediction device 120 and the composition optimization device 130 determine whether to perform a task of predicting phase fraction or a task of exploring material composition.

[0124] If it is determined in step S303 that the task of predicting the phase fraction is to be executed, the process proceeds to step S304, and the prediction device 120 accepts the selection of the material composition information. Furthermore, in step S305, the prediction device 120 executes the prediction process on the selected material composition information.

[0125] On the other hand, when it is determined in step S303 that the task of exploring the material composition is to be performed, the process proceeds to step S306, and the composition optimization device 130 accepts the setting of the target phase fraction. In addition, in step S307, the composition optimization device 130 performs combinatorial optimization processing to explore the material composition as the accepted target phase fraction.

[0126] <Composition of training data, input data, and correct answer data>

[0127] Next, a specific example of the structure of training data and input data and correct answer data included in the training data will be described. Figure 4 This is a specific example showing the structure of training data, input data, and correct answer data.

[0128] like Figure 4 As shown, the training data 400 includes "input data" and "correct answer data" as information items.

[0129] "Material composition information 1", "Material composition information 2", ..., etc., which are different material composition information, are stored in "Input data". Figure 4 In the figure, reference numeral 401 indicates a specific example of "material composition information 1", and shows the weight % of each additive constituting the known aluminum alloy standard labeled "6013".

[0130] Similarly, reference numeral 402 indicates a specific example of "material composition information 2", and shows the weight % of each additive constituting "6060" which is a symbol of a known aluminum alloy specification.

[0131] In the “correct answer data”, the predetermined temperature intervals ( Figure 4 In the example, it is the phase fraction at each temperature of 100°C to 700°C, that is, "phase fraction 1", "phase fraction 2", ..., etc. Figure 4, reference numeral 411 represents a specific example of “phase fraction 1”, and reference numeral 412 represents a specific example of “phase fraction 2”, wherein the horizontal axis represents temperature and the vertical axis represents phase fraction.

[0132] Here, reference numerals 411 and 412 are results obtained by running a high-cost simulator with high prediction accuracy for a long period of time, and in the present embodiment, these results are used as correct answer data.

[0133] It should be noted that the simulator referred to here refers to software that calculates the phase fraction at each temperature within a predetermined temperature range by thermodynamic equilibrium calculation, such as software such as CaTCalc (registered trademark) and MatCalc (registered trademark). Alternatively, the simulator may also include software such as Termosuite (registered trademark), FactStage (registered trademark), and Pandat (registered trademark). Alternatively, the simulator may also include software such as MALT2 (registered trademark), Thermo-Calc (registered trademark), and OpenCalphad (registered trademark).

[0134] <Details of each unit of the training device>

[0135] Next, each unit of the training device 110 (the training data generating unit 111 and the training unit 112) will be described in detail.

[0136] (1) Functional structure of training data generation unit

[0137] First, the functional configuration of the training data generating unit 111 will be described. Figure 5 This is a diagram showing an example of the functional structure of the training data generating unit.

[0138] like Figure 5 As shown, the training data generating unit 111 includes an element information input unit 501 , a combination determining unit 502 , a simulation unit 503 , and a storage control unit 504 .

[0139] The element information input unit 501 receives input of, for example, the type of additional element added when forming a specific alloy. In addition, the element information input unit 501 receives input of the upper limit and lower limit of the addition amount (weight %) of each additional element received as input.

[0140] The combination determination unit 502 determines multiple combinations of the addition amount of each additional element by randomly selecting the addition amount based on the constraints of the upper limit value and the lower limit value for each additional element received as input. In addition, the combination determination unit 502 notifies the simulation unit 503 of the material composition information represented by the determined multiple combinations.

[0141] The simulation unit 503 calculates the phase fraction at each temperature in the predetermined temperature range by operating the above-mentioned simulator for each of the plurality of material composition information notified by the combination determination unit 502. In addition, the simulation unit 503 notifies the storage control unit 504 of the plurality of material composition information and the corresponding phase fraction at each temperature in the predetermined temperature range.

[0142] The storage control unit 504 generates training data 400 using the plurality of material composition information notified by the simulation unit 503 as “input data” and the phase fractions at each temperature in the corresponding predetermined temperature range as “correct answer data”, and stores the training data in the training data storage unit 113 .

[0143] (2) Functional structure of training units

[0144] Next, the functional structure of the training unit 112 will be described. Figure 6 This is a diagram showing an example of the functional structure of a training unit.

[0145] like Figure 6 As shown, the training unit 112 has a prediction model and a loss function calculation unit 603 .

[0146] As described above, Seq2Seq or Transformer with Attention mechanism is applicable to the prediction model, including encoder unit 601 and decoder unit 602.

[0147] Encoder section 601 outputs a feature quantity by inputting “material composition information 1”, “material composition information 2”, etc. stored in “input data” of training data 400 .

[0148] The decoder unit 602 is input with the feature quantity output by the encoder unit 601, and sequentially outputs the output data. More specifically, the decoder unit 602 reads out the "phase fraction 1", "phase fraction 2", etc. stored in the "correct data" of the training data 400. Then, the decoder unit 602 uses the phase fraction up to the i-th temperature (that is, the correct data up to the i-th temperature) to output the output data of the i+1-th temperature. Thus, the decoder unit 602 can sequentially output the output data of the phase fraction at each temperature within the predetermined temperature range based on, for example, the feature quantity output from the encoder unit 601 when the "material composition information 1" is input, and the "correct data" of the "phase fraction 1". Similarly, the decoder unit 602 can sequentially output the output data of the phase fraction at each temperature within the predetermined temperature range based on, for example, the feature quantity output from the encoder unit 601 when the "material composition information 2" is input, and the "correct data" of the "phase fraction 2".

[0149] It should be noted that the Softmax function is used as the output layer of the decoder unit 602. Therefore, the multiple values ​​(i.e., the values ​​of each phase) included in the output data of the decoder unit 602 range from 0 to 1 and the sum of the multiple values ​​(i.e., the sum of the values ​​of each phase) is "1".

[0150] In this way, by applying the Softmax function to the output layer of the decoder unit 602, output data corresponding to the phase fraction at each temperature within a predetermined temperature range can be output in sequence without performing additional calculations.

[0151] The loss function calculation unit 603 reads out “phase fraction 1”, “phase fraction 2”, etc. stored in the “correct answer data” of the training data 400 , and compares them with the output data at the i+1th temperature output by the decoder unit 602 .

[0152] In addition, the loss function calculation unit 603 calculates multiple losses when performing the comparison, and performs weighted addition on the calculated multiple losses to calculate the comprehensive loss. Furthermore, the loss function calculation unit 603 updates the model parameters of the encoder unit 601 and the decoder unit 602 according to the calculated comprehensive loss. Thus, a trained prediction model (including a trained encoder unit and a trained decoder unit) is generated.

[0153] (3) Operation Example 1 of Training Unit

[0154] Next, an operation example of the training unit 112 (mainly an operation example of the prediction model) will be described. Figure 7 FIG. 1 is a first diagram showing an operation example of the training unit.

[0155] in, Figure 7 (a) shows a situation where “input data” of training data 400 is read out and input to encoder section 601 , and feature quantities are output from encoder section 601 . Figure 7 (a) shows a situation when a start signal is input to the decoder unit 602 so that the decoder unit 602 outputs output data. Figure 7 In the example of (a), the output data output by the decoder unit 602 is equivalent to the phase fraction at 700°C. It should be noted that the start signal is a signal for starting the operation of the decoder unit 602. In this embodiment, an arbitrary value (such as "000000" or the like) different from the output data output from the decoder unit 602 is input to the decoder unit 602 as the start signal. The output data output from the decoder unit 602 (here, the output data equivalent to the phase fraction at 700°C) is notified to the loss function calculation unit 603.

[0156] Figure 7(b) shows a situation where the feature quantity output from the encoder unit 601 is input to the decoder unit 602 and the phase fraction at 700°C (correct data at 700°C) is input to the decoder unit 602, thereby outputting output data from the decoder unit 602. Figure 7 In the example (b), the output data output by the decoder unit 602 is equivalent to the phase fraction at 690° C. It should be noted that the output data output from the decoder unit 602 (here, the output data equivalent to the phase fraction at 690° C.) is notified to the loss function calculation unit 603.

[0157] Figure 7 (c) shows a situation where the feature quantity output from the encoder unit 601 is input to the decoder unit 602 and the phase ratio up to 690°C (correct data up to 690°C) is input to the decoder unit 602 so that output data is output from the decoder unit 602. However, in Figure 7 Due to space limitations, only the correct answer data at 690°C are recorded in (c) (in fact, the correct answer data at 700°C and the correct answer data at 690°C are weighted and input into the decoder unit 602).

[0158] In addition, Figure 7 In the example of (c), the output data output by the decoder unit 602 is equivalent to the phase fraction at 680° C. It should be noted that the output data output from the decoder unit 602 (here, the output data equivalent to the phase fraction at 680° C.) is notified to the loss function calculation unit 603 .

[0159] Figure 7 (d) shows a situation where the feature quantity output from the encoder unit 601 is input to the decoder unit 602 and the phase ratio up to 120°C (correct solution data up to 120°C) is input to the decoder unit 602 so that output data is output from the decoder unit 602. However, in Figure 7 Due to space limitations, only the correct answer data at 120°C are recorded in (d) (in fact, the correct answer data at 700°C to 120°C are weighted and input into the decoder unit 602).

[0160] In addition, Figure 7 In the example of (d), the output data output by decoder section 602 corresponds to the phase fraction at 110° C. The output data output from decoder section 602 (here, the output data corresponding to the phase fraction at 110° C.) is notified to loss function calculation section 603 .

[0161] Figure 7(e) shows a case where the feature quantity output from the encoder unit 601 is input to the decoder unit 602 and the phase ratio up to 110°C (correct solution data up to 110°C) is input to the decoder unit 602, thereby outputting output data from the decoder unit 602. Figure 7 Due to space limitations, only the correct answer data at 110°C are recorded in (e) (actually, the correct answer data at 700°C to 110°C are weighted and input into the decoder unit 602).

[0162] exist Figure 7 In the example of (e), the output data output by the decoder unit 602 is equivalent to the phase fraction at 100° C. It should be noted that the output data output from the decoder unit 602 (here, the output data equivalent to the phase fraction at 100° C.) is notified to the loss function calculation unit 603.

[0163] (4) Operation Example 2 of Training Unit

[0164] Next, an operation example of the training unit 112 (mainly an operation example of the loss function calculation unit 603) is described. Figure 8 FIG. 2 is a second diagram showing an operation example of the training unit. Figure 8 As shown, the loss function calculation unit 603 reads out the “correct answer data” of the training data 400 . Figure 8 The example shows a case where the phase fraction at each temperature in the following predetermined temperature range included in "Phase Fraction 1" is read out:

[0165] Phase fraction at 700°C (correct data),

[0166] Phase fraction at 690°C (correct data), ...

[0167] Phase fraction at 100°C (correct data).

[0168] also, Figure 8 The example shows the following situation: based on the feature quantity output from the encoder unit 601 by inputting "material composition information 1", the output data is sequentially output by the decoder unit 602, and as a result, in the loss function calculation unit 603, the following phase fraction is maintained:

[0169] Phase fraction at 700°C (output data),

[0170] Phase fraction at 690°C (output data), ...

[0171] Phase fraction at 100°C (output data).

[0172] also, Figure 8The example of FIG. 6 shows a case where the loss function calculation unit 603 calculates the comprehensive loss by comparing the phase fraction (correct answer data) at each temperature of 100°C to 700°C with the phase fraction (output data) at each temperature of 100°C to 700°C. In addition, Figure 8 The example shows a situation where the loss function calculation unit 603 updates the model parameters of the encoder unit 601 and the decoder unit 602 according to the calculated comprehensive loss. It should be noted that the functional structure of the loss function calculation unit 603 for calculating the comprehensive loss is described in detail below.

[0173] (5) Details of the functional structure of the loss function calculation unit

[0174] Fig. 9 is a diagram showing an example of a method for calculating the loss performed by the loss function calculation unit. Fig. 9 As shown, the loss function calculation unit 603 has

[0175] Phase fraction error calculation unit 901 at generation / disappearance temperature,

[0176] Phase fraction error calculation unit 902,

[0177] Phase ratio error (logarithmic value) calculation unit 903,

[0178] Phase ratio error (difference value) calculation unit 904,

[0179] Cross entropy error calculation unit 905,

[0180] A weighted addition unit 906,

[0181] As a function for calculating multiple losses.

[0182] The phase fraction error calculation unit 901 at the generation / disappearance temperature compares the phase fraction (correct answer data) at each temperature of 100°C to 700°C with the phase fraction (output data) at each temperature of 100°C to 700°C for the temperature at which the curve of each phase fraction in the correct answer data of the training data is 0. Thus, the phase fraction error calculation unit 901 at the generation / disappearance temperature calculates the phase fraction error at the generation or disappearance temperature of each phase determined based on the training data.

[0183] Phase fraction and

[0184] Phase fraction determined based on training data

[0185] The errors are added in all phases and the addition result (the first addition result) is output.

[0186] The phase fraction error calculation unit 902 compares the phase fraction (correct answer data) at each temperature of 100° C. to 700° C. with the phase fraction (output data) at each temperature of 100° C. to 700° C. Thus, the phase fraction error calculation unit 902 adds the errors of the phase fraction at each temperature in a predetermined temperature range and outputs the addition result (second addition result).

[0187] The phase fraction error (logarithmic value) calculation unit 903 compares the phase fraction (correct answer data) at each temperature of 100° C. to 700° C. with the phase fraction (output data) at each temperature of 100° C. to 700° C. Thus, the phase fraction error (logarithmic value) calculation unit 903 adds the errors of the logarithmic values ​​of the phase fraction at each temperature in the predetermined temperature range in the predetermined temperature range, and outputs the addition result (third addition result).

[0188] The phase fraction error (difference value) calculation unit 904 compares the phase fraction (correct answer data) at each temperature of 100° C. to 700° C. with the phase fraction (output data) at each temperature of 100° C. to 700° C. Thus, the phase fraction error (difference value) calculation unit 904 adds the errors of the difference values ​​of the phase fractions between adjacent temperatures in a predetermined temperature range and outputs the addition result (the fourth addition result).

[0189] The cross entropy error calculation unit 905 compares the phase fraction (correct answer data) at each temperature of 100° C. to 700° C. with the phase fraction (output data) at each temperature of 100° C. to 700° C. Thus, the cross entropy error calculation unit 905 adds the errors of the ratio of the phase fractions at each temperature in the predetermined temperature interval in the predetermined temperature interval, and outputs the addition result (the fifth addition result).

[0190] The weighted addition unit 906 calculates the comprehensive loss by performing weighted addition on the addition results (the first addition result to the fifth addition result) respectively output from the phase fraction error calculation unit 901 to the cross entropy error calculation unit 905 of the generation / disappearance temperature.

[0191] Thus, by adopting a structure of calculating multiple losses and weighted addition thereof, the loss function calculation unit 603 can process the loss between the output data and the correct answer data in multiple ways. As a result, in the training unit 112, when training the prediction model, the model parameters can be appropriately updated.

[0192] <Flow of training process>

[0193] Next, the flow of the training process by the training device 110 will be described. Fig.10 It is a flowchart which shows the flow of a training process.

[0194] In step S1001 , the training data generating unit 111 receives input of the type of additional elements added when generating a specific alloy and the upper limit and lower limit of the addition amount (weight %) of each additional element as element information.

[0195] In step S1002 , the training data generating unit 111 determines a plurality of combinations of the amount of addition of each additional element by randomly selecting the amount of addition of the additional element based on the constraints of the upper limit value and the lower limit value.

[0196] In step S1003 , the training data generating unit 111 operates the simulator for each of the material composition information represented by the determined plurality of combinations, thereby calculating the phase fraction at each temperature within the predetermined temperature range.

[0197] In step S1004 , training data generating section 111 generates training data and stores the training data storing section 113 .

[0198] In step S1005 , the training unit 112 inputs a start signal to the decoder unit 602 .

[0199] In step S1006 , the training unit 112 reads out the “input data” of the training data, and inputs it into the prediction model.

[0200] In step S1007 , the training unit 112 sequentially outputs (i+1) th output data by sequentially inputting the phase fractions up to the i th temperature of the “correct answer data” of the training data (ie, the correct answer data up to the i th temperature).

[0201] In step S1008, the training unit 112 determines whether all output data corresponding to the phase fraction at each temperature in the predetermined temperature range has been output. If it is determined in step S1008 that there is a temperature at which output data corresponding to the phase fraction has not been output (in the case of "No" in step S1008), the process returns to step S1007.

[0202] On the other hand, when it is determined in step S1008 that all output data corresponding to the phase fractions at the respective temperatures within the predetermined temperature interval have been output (in the case of YES in step S1008 ), the process proceeds to step S1009 .

[0203] In step S1009 , the training unit 112 reads out the “correct answer data” of the training data, and compares it with the output data.

[0204] In step S1010, the training unit 112 calculates multiple types of losses and performs weighted addition on the calculated multiple losses to calculate the comprehensive loss. In addition, the training unit 112 updates the model parameters of the encoder unit 601 and the decoder unit 602 according to the calculated comprehensive loss.

[0205] In step S1011, the training unit 112 determines whether to continue the training process. If it is determined in step S1011 that the training process is to be continued (in the case of "yes" in step S1011), the process returns to step S1006 and the same process as above is performed by reading the next "input data" of the training data.

[0206] On the other hand, when it is determined in step S1011 that the training process is to be ended (in the case of NO in step S1011 ), the process proceeds to step S1012 .

[0207] In step S1012, the training unit 112 saves the trained encoder unit and the trained decoder unit as trained prediction models in the prediction device 120 and the composition optimization device 130.

[0208] <Details of each unit of the prediction device>

[0209] Next, each unit of the prediction device 120 (here, the prediction unit 122) will be described in detail.

[0210] (1) Functional structure of prediction unit

[0211] First, the functional structure of the prediction unit 122 will be described. Fig.11 This is a diagram showing an example of the functional structure of a prediction unit.

[0212] like Fig.11 As shown, the prediction unit 122 includes a trained encoder unit 1101 and a trained decoder unit 1102 that have been trained by the training unit 112. The trained encoder unit 1101 and the trained decoder unit 1102 form a trained prediction model.

[0213] The trained encoder unit 1101 calculates a feature quantity by inputting the material composition information (material composition information of the prediction target material) notified by the material composition input unit 121 , and outputs the calculated feature quantity to the trained decoder unit 1102 .

[0214] The trained decoder unit 1102 sequentially outputs prediction data by inputting the feature quantity output from the trained encoder unit 1101. Specifically, the trained decoder unit 1102 outputs prediction data at the i+1th temperature using the prediction data up to the i-th temperature. Thus, the trained decoder unit 1102 can predict the phase fraction at each temperature within a predetermined temperature range based on the feature quantity output from the trained encoder unit 1101 by inputting the material composition information of the prediction target material.

[0215] (2) Example of operation of prediction unit

[0216] Next, an operation example of the prediction unit 122 is described. Fig.12 It is a diagram showing an operation example of the prediction unit. Fig.12 (a) shows a case where the feature quantity is output from the trained encoder unit 1101 by inputting the material composition information of the prediction target material into the trained encoder unit 1101. In addition, Fig.12 (a) shows a case where the feature amount output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and a start signal is input to the trained decoder unit 1102, thereby outputting prediction data (phase fraction at 700°C).

[0217] Fig.12 (b) shows a situation where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and the prediction data (phase fraction at 700° C.) is input to the trained decoder unit 1102. Thus, the trained decoder unit 1102 outputs the prediction data (phase fraction at 690° C.).

[0218] Fig.12 (c) shows a case where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and the prediction data (up to the phase fraction of 690° C.) is input to the trained decoder unit 1102. Fig.12 In (c), due to space limitations, only the phase fraction at 690°C is recorded (actually, the phase fraction at 700°C and the phase fraction at 690°C are weighted and input into the trained decoder unit 1102). Thus, the trained decoder unit 1102 outputs predicted data (phase fraction at 680°C).

[0219] Fig.12 (d) shows a case where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and the prediction data (up to the phase fraction of 120°C) is input to the trained decoder unit 1102. Fig.12In (d), due to space limitations, only the phase fraction at 120°C is recorded (actually, the phase fractions at 700°C to 120°C are weighted and input to the trained decoder unit 1102.) Thus, the trained decoder unit 1102 outputs predicted data (phase fraction at 110°C).

[0220] Fig.12 (e) shows a case where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and the prediction data (up to the phase fraction of 110°C) is input to the trained decoder unit 1102. Fig.12 In (e), due to space limitations, only the phase fraction at 110°C is recorded (actually, the phase fractions at 700°C to 110°C are weighted and input to the trained decoder unit 1102.). As a result, the trained decoder unit 1102 outputs the predicted data (the phase fraction at 100°C). It should be noted that Fig.12 In the example of (e), the predicted data (phase fraction at 100° C.) output from the trained decoder unit 1102 by inputting the feature quantity and the predicted data (phase fraction up to 110° C.) is also shown.

[0221] <Flow of prediction processing>

[0222] Next, the flow of prediction processing by the prediction device 120 will be described. Fig.13 It is a flowchart which shows the flow of prediction processing.

[0223] In step S1301 , the material composition input unit 121 receives input of material composition information on a prediction target material.

[0224] In step S1302 , the prediction unit 122 receives a start signal and inputs the input material composition information into the trained prediction model, thereby predicting the phase fraction at each temperature within a predetermined temperature range.

[0225] In step S1303 , the display unit 123 displays the predicted phase fraction at each temperature.

[0226] <Prediction accuracy>

[0227] Next, the prediction accuracy of the phase fraction at each temperature within the predetermined temperature range predicted by the prediction unit 122 will be described. Fig.14 It is a graph for explaining the prediction accuracy. Fig.14 (a) to (c) of FIG. 1 show the predicted data when the weight percentage of each added element constituting the known aluminum alloy specification "6013" is input as the material composition information of the predicted object material. It should be noted that in this embodiment, when evaluating the prediction accuracy, the weight percentage of each added element constituting the known aluminum alloy specification "6013" is used. Figure 4The correct answer data is shown in the reference numeral 411. In calculating the loss with the correct answer data, the mean square logarithmic error (MSLE: Mean Squared Logarithmic Error) loss is used.

[0228] Fig.14 (a) shows the prediction data when a fully connected multilayer neural network is applied to the prediction model as a comparative example. Fig.14 In the case of (a), the MSLE loss compared with the correct solution data is 3.79×10 -4 .

[0229] Fig.14 (b) shows the prediction data when Seq2Seq with Attention mechanism is applied to the prediction model. Fig.14 In the case of (b), the MSLE loss compared with the correct solution data is 5.1×10 -5 .

[0230] Fig.14 (c) shows the prediction data when Transformer is applied to the prediction model. Fig.14 In the case of (c), the MSLE loss compared with the correct solution data is 2.45×10 -5 .

[0231] As described above, when a fully connected multilayer neural network is adopted as a prediction model for realizing phase fraction calculation at low cost, the estimation accuracy is low in some temperature regions (eg, 400° C. to 600° C.).

[0232] On the other hand, when Seq2Seq with Attention mechanism is used as a prediction model to realize phase fraction calculation at low cost, the loss can be greatly improved. In addition, when Transformer is used, the phase fraction at the same level as the correct answer data can be reproduced.

[0233] That is, when predicting the phase fraction in a predetermined temperature range based on the material composition information,

[0234] Using a recursive network, the phase fraction at each temperature is processed as time series data.

[0235] Use multiple loss functions and handle losses in multiple ways,

[0236] Can improve prediction accuracy.

[0237] <Details of processing in composition optimization device>

[0238] Next, the details of the processing in the composition optimization device 130 are described.

[0239] (1) Details of the forward propagation process in the composition optimization device

[0240] First, the details of the forward propagation processing in the composition optimization device 130 (mainly the details of the processing in the prediction unit 132) are explained. Fig.15 FIG1 is the first figure showing an operation example of the forward propagation processing in the composition optimization device.

[0241] like Fig.15 As shown, like the prediction unit 122, the prediction unit 132 includes a trained encoder unit 1101 and a trained decoder unit 1102 that have been trained by the training unit 112. The trained encoder unit 1101 and the trained decoder unit 1102 form a trained prediction model.

[0242] In the forward propagation processing of the composition optimization device 130 , the trained encoder unit 1101 calculates the feature quantity by inputting the default material composition information notified by the material composition input unit 131 , and outputs the calculated feature quantity to the trained decoder unit 1102 .

[0243] In addition, in the forward propagation processing of the composition optimization device 130, the trained decoder unit 1102 sequentially outputs prediction data by inputting the feature quantity output from the trained encoder unit 1101. Specifically, the trained decoder unit 1102 uses the prediction data up to the i-th temperature to output the prediction data at the i+1-th temperature. Thus, the trained decoder unit 1102 is able to predict the phase fraction at each temperature within a predetermined temperature range based on the feature quantity output from the trained encoder unit 1101 by inputting the default material composition information. It should be noted that in the forward propagation processing of the composition optimization device 130, the prediction data output by the trained decoder unit 1102 is input to the loss function calculation unit 133.

[0244] (2) Details of the forward propagation process in the composition optimization device (Part 2)

[0245] Next, further details of the forward propagation processing in the composition optimization device 130 (mainly further details of the processing in the prediction unit 132) are described. Fig.16 FIG2 is a second diagram showing an operation example of the forward propagation processing in the composition optimization device. Fig.16 (a) shows a situation where a feature amount is output from the trained encoder unit 1101 by inputting default material composition information to the trained encoder unit 1101. Fig.16(a) shows a case where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and a start signal is input to the trained decoder unit 1102 to output prediction data (phase fraction at 700° C.). Fig.16 (a) shows a situation where the output prediction data (phase fraction at 700° C.) is input to the loss function calculation unit 133 .

[0246] Fig.16 (b) shows a situation where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and the prediction data (phase fraction at 700°C) is input to the trained decoder unit 1102. As a result, the trained decoder unit 1102 outputs the prediction data (phase fraction at 690°C). It should be noted that the prediction data (phase fraction at 690°C) output from the trained decoder unit 1102 is input to the loss function calculation unit 133.

[0247] Fig.16 (c) shows a case where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and the prediction data (up to the phase fraction of 690° C.) is input to the trained decoder unit 1102. Fig.16 In (c), due to space limitations, only the phase fraction at 690°C is recorded (actually, the phase fraction at 700°C and the phase fraction at 690°C are weighted and input to the trained decoder unit 1102). As a result, the trained decoder unit 1102 outputs predicted data (phase fraction at 680°C). It should be noted that the predicted data (phase fraction at 680°C) output from the trained decoder unit 1102 is input to the loss function calculation unit 133.

[0248] Fig.16 (d) shows a case where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and the prediction data (up to the phase fraction of 120°C) is input to the trained decoder unit 1102. Fig.16 In (d), due to space limitations, only the phase fraction at 120°C is recorded (actually, the phase fractions at 700°C to 120°C are weighted and input to the trained decoder unit 1102). As a result, the trained decoder unit 1102 outputs predicted data (phase fraction at 110°C). It should be noted that the predicted data (phase fraction at 120°C) output from the trained decoder unit 1102 is input to the loss function calculation unit 133.

[0249] Fig.16(e) shows a case where the feature quantity output from the trained encoder unit 1101 is input to the trained decoder unit 1102 and the prediction data (up to the phase fraction of 110°C) is input to the trained decoder unit 1102. Fig.16 In (e), due to space limitations, only the phase fraction at 110°C is recorded (actually, the phase fractions at 700°C to 110°C are weighted and input to the trained decoder unit 1102). As a result, the trained decoder unit 1102 outputs the predicted data (the phase fraction at 100°C). It should be noted that Fig.16 In the example (e), the situation in which the predicted data (phase fraction at 100°C) is output from the trained decoder unit 1102 by inputting the feature value and the predicted data (phase fraction up to 110°C) and is input to the loss function calculation unit 133 is also shown.

[0250] (3) Details of the forward propagation process in the composition optimization device (Part 3)

[0251] Next, the details of the forward propagation processing in the composition optimization device 130 (mainly the details of the processing in the loss function calculation unit 133) are explained. Fig.17 FIG3 is a third diagram showing an operation example of the forward propagation processing in the composition optimization device.

[0252] like Fig.17 As shown, in the forward propagation process in the composition optimization device 130, the target phase fraction is preset in the loss function calculation unit 133. Fig.17 The example shows the preset

[0253] Phase fraction at 700°C (target data),

[0254] Phase fraction at 690°C (target data), ...

[0255] Phase fraction at 100°C (target data),

[0256] As the target phase fraction.

[0257] In addition, as described above, in the forward propagation process of the composition optimization device 130, the loss function calculation unit 133 is input

[0258] Phase fraction at 700°C (predicted data),

[0259] ·Phase fraction at 690℃ (predicted data),……

[0260] Phase fraction at 100°C (predicted data),

[0261] As the phase fraction at each temperature within a predetermined temperature interval output by the trained decoder unit 1102 .

[0262] In addition, in the forward propagation processing in the composition optimization device 130, the loss function calculation unit 133 calculates the comprehensive loss based on the phase fraction (target data, predicted data) at each temperature within a predetermined temperature range. Fig.17 The example shows a case where the comprehensive loss is calculated by comparing the target phase fraction (target data) at each temperature of 100°C to 700°C with the predicted phase fraction (predicted data) at each temperature of 100°C to 700°C.

[0263] (4) Details of back propagation processing in the composition optimization device

[0264] Next, the details of the back-propagation processing in the composition optimization device 130 (mainly the details of the processing in the loss function calculation unit 133, the prediction unit 132, and the update unit 134) are explained. Fig.18 It is a diagram showing an operation example of the back-propagation processing in the composition optimization device.

[0265] like Fig.18 As shown, in the back propagation process in the composition optimization device 130, the loss function calculation unit 133 back propagates the comprehensive loss corresponding to the default material composition information in a state where the model parameters of the trained prediction model of the prediction unit 132 are fixed. Thus, the back propagation result is output from the trained prediction model of the prediction unit 132 and input to the updating unit 134.

[0266] In addition, if Fig.18 As shown, in the back propagation process of the composition optimization device 130, the updating unit 134 updates the material composition information based on the back propagation result output from the prediction unit 132 and the previous material composition information (here, the default material composition information). Thus, the updating unit 134 generates the updated material composition information.

[0267] (5) Details of the forward propagation process in the composition optimization device (Part 4)

[0268] Next, the details of the forward propagation processing in the composition optimization device 130 (mainly the details of the processing in the prediction unit 132) are explained. Fig.19 FIG4 is a fourth diagram showing an operation example of the forward propagation process in the composition optimization device. Fig.15 The difference is that in Fig.19 In the case of, the updated material composition information generated by the updating unit 134 is input to the trained encoder unit 1101. The forward propagation processing and the reverse propagation processing after the updated material composition information is input to the trained encoder unit 1101 are the same as the reference Figures 16 to 18The forward propagation process and the reverse propagation process described are the same, so their description is omitted here.

[0269] <Flow of composition optimization processing>

[0270] Next, the flow of the composition optimization processing performed by the composition optimization device 130 is described. Fig. 20 It is a flowchart which shows the flow of composition optimization processing.

[0271] In step S2001 , the loss function calculation unit 133 accepts the setting of the target phase fraction.

[0272] In step S2002 , the material composition input unit 131 accepts input of default material composition information.

[0273] In step S2003 , the prediction unit 132 receives the start signal and inputs the input material composition information into the trained prediction model, thereby outputting the predicted phase fraction (prediction data) at each temperature within the predetermined temperature range.

[0274] In step S2004, the loss function calculation unit 133 compares the predicted phase fraction (prediction data) output from the trained prediction model with the target phase fraction (target data).

[0275] In step S2005, the loss function calculation unit 133 calculates the comprehensive loss as a result of the comparison.

[0276] In step S2006, the loss function calculation unit 133 determines whether the comprehensive loss satisfies a predetermined condition. If it is determined in step S2006 that the predetermined condition is not satisfied (in the case of "No" in step S2006), the process proceeds to step S2007.

[0277] In step S2007, the prediction unit 132 back-propagates the calculated comprehensive loss while fixing the model parameters of the trained prediction model, and outputs the back-propagation result.

[0278] In step S2008, the updating unit 134 updates the previous material composition information based on the outputted back propagation result. It should be noted that the process of updating the previous material composition information will be described in detail later.

[0279] In step S2009 , the material composition input unit 131 accepts input of updated material composition information, and the process returns to step S2003 .

[0280] On the other hand, when it is determined in step S2006 that the predetermined condition is satisfied (in the case of "YES" in step S2006), the material composition information when it is determined that the predetermined condition is satisfied is used as the search result, and the composition optimization process is ended.

[0281] <Details of material composition update process>

[0282] Next, the process of updating the previous material composition information (step S2008) will be described in detail. Fig.21 This is a flowchart showing the flow of material composition update processing.

[0283] In step S2101 , the updating unit 134 obtains the back-propagation result L output from the trained prediction model of the prediction unit 132 .

[0284] In step S2102 , the updating unit 134 inputs “1” to the counter i.

[0285] In step S2103, the updating unit 134 partially differentiates the ratios X_i of the previous material composition information using the back propagation result L. It should be noted that the result of partially differentiating the ratios X_i of the respective components in the back propagation result L is a value indicating whether the error increases or decreases when the ratios X_i of the corresponding components are increased. The updating unit 134 increases or decreases the ratios X_i of the respective components in a manner that reduces the comprehensive loss.

[0286] In step S2104, if the result of partial differentiation in step S2103 is a negative value, the updating unit 134 increases the ratio X_i of the corresponding composition. It should be noted that the result of partial differentiation being a negative value indicates that the error decreases when the ratio X_i of the corresponding composition increases.

[0287] In step S2105, if the result of the partial derivative in step S2103 is a positive value, the updating unit 134 reduces the ratio Xi of the corresponding composition. The positive value of the partial derivative result indicates that the error increases when the ratio Xi of the corresponding composition increases.

[0288] In step S2106 , the updating unit 134 determines whether all compositions included in the previous material composition information have been updated.

[0289] When it is determined in step S2106 that there is a non-updated composition (in the case of NO in step S2106 ), the process proceeds to step S2107 .

[0290] In step S2107, the updating unit 134 increments the counter i and then returns to step S2104.

[0291] On the other hand, when it is determined in step S2106 that all compositions have been updated (in the case of YES in step S2106 ), the material composition update processing is terminated.

[0292] <Results of composition optimization processing>

[0293] Next, the processing results of the composition optimization processing performed by the composition optimization device 130 are described. Fig. 22 FIG1 is a first diagram showing an example of a processing result of the composition optimization processing.

[0294] exist Fig. 22 , reference numeral 2200 represents an example of a phase fraction corresponding to the default material composition information input to the prediction unit 132.

[0295] In addition, Fig. 22 , reference numeral 2210 represents an example of a target phase fraction set in the loss function calculation unit 133 .

[0296] In addition, Fig. 22 , reference numeral 2211 indicates a predicted phase fraction when the composition optimization device 130 of the present embodiment performs composition optimization processing on the target phase fraction in such a manner that the comprehensive loss satisfies a predetermined condition. Specifically, the predicted phase fraction outputted by inputting the updated material composition information input to the prediction unit 132 after the material composition information is updated 515 times to the prediction unit 132 is shown.

[0297] On the other hand, Fig. 22 , as a comparative example, reference numeral 2212 indicates the predicted phase fraction when only RMSLE is calculated to perform composition optimization processing when the loss function calculation unit 133 calculates the comprehensive loss. Specifically, the predicted phase fraction outputted by inputting the updated material composition information input to the prediction unit 132 after updating the material composition information input to the prediction unit 132 800 times is shown. It should be noted that RMSLE is the abbreviation of Root Mean Squared Logarithmic Error, which refers to the square root of the mean square logarithmic error.

[0298] From the comparison between reference numerals 2210 and 2211 , it can be seen that according to the composition optimization device 130 of this embodiment, at the stage where the material composition information is updated 515 times, a material composition having a predicted phase fraction similar to the target phase fraction has been explored.

[0299] On the other hand, as can be seen from the comparison between the reference mark 2210 and the reference table 2212, if the loss function calculation unit 133 does not calculate the appropriate loss, even if the material composition information is updated 800 times, the material composition with a predicted phase fraction similar to the target phase fraction cannot be explored.

[0300] Therefore, if appropriate losses are not calculated when calculating the comprehensive losses, it takes time to search for material composition information that satisfies the predetermined conditions for the target phase fraction. On the contrary, if appropriate losses are calculated when calculating the comprehensive losses, the time spent on searching for material composition information that satisfies the predetermined conditions for the target phase fraction can be reduced.

[0301] Summary

[0302] As is clear from the above description, the material development support system 100 according to the first embodiment

[0303] · generating a trained model by performing training using training data that associates material composition information of a training target material with phase fractions of the training target material at each temperature within a predetermined temperature range,

[0304] Predicting the phase fractions of a material with a predetermined material composition at each temperature within a predetermined temperature interval by inputting a default material composition into the generated trained model,

[0305] Updating the predetermined material composition input to the trained model by back-propagating the combined loss calculated based on the target phase fraction (target data) and the predicted phase fraction (predicted data),

[0306] The process of outputting the predicted phase fraction (prediction data) based on the updated material composition and the process of updating the default material composition by the comprehensive loss calculated by back propagation are repeated until the comprehensive loss satisfies a predetermined condition.

[0307] As described above, in the material development support system 100 according to the first embodiment, when searching for a material composition corresponding to a target phase fraction, the material composition is changed by backpropagating comprehensive losses instead of solving a large number of forward problems while comprehensively changing the material composition.

[0308] Thus, according to the material development support system 100 of the first embodiment, it is possible to efficiently search for material composition information corresponding to a target phase fraction.

[0309] [Second embodiment]

[0310] In the first embodiment described above, the previous material composition information is freely updated using the back propagation result output in the back propagation process. On the other hand, since the material composition information is information indicating the ratio of each composition, the material composition information must always be a positive value and the total value must be 100% (1.0). Therefore, in the second embodiment, the updated material composition information is corrected so that the updated material composition information is always a positive value and the total value is 100% (1.0). The second embodiment will be described below, centering on the differences from the first embodiment described above.

[0311] <Functional structure of composition optimization device>

[0312] First, the functional structure of the composition optimization device according to the second embodiment will be described. Fig.23 FIG. 1 is another example of the functional structure of the optimization device. Figure 1 The functional configuration of the composition optimization device 130 described above is different in that the composition optimization device 2300 of the second embodiment includes a constraint unit 2301 .

[0313] The constraint unit 2301 corrects the updated material composition information so that the total value of the ratio of each component is 100%. For example, when the material is an aluminum alloy, the ratio of the aluminum element is calculated by subtracting the sum of the ratios of each element other than the aluminum element from 100%. Thus, in the constraint unit 2301, the total value of the ratio of each component can be made 100% for the updated material composition information.

[0314] Furthermore, if the composition ratio X_i becomes negative due to the reduction of the composition ratio X_i in the updating unit 134, the constraint unit 2301 corrects the composition ratio X_i to zero to avoid the composition ratio X_i becoming negative. That is, the constraint unit 2301 corrects the updated material composition information so that the ratio of each composition is 0% or more.

[0315] It should be noted that the difference between the total value of the ratio of each component and 100% may not be adjusted, but may be adjusted, for example, by uniformly reducing the ratio of the components other than the composition ratio X_i as the updated material composition information. Alternatively, the adjustment may be made by reducing the composition with the maximum ratio so that the total value of the ratio of each component is 100%.

[0316] Summary

[0317] As is clear from the above description, in the material development support system 100 according to the second embodiment, the composition optimization device 130 has a constraint unit 2301 to prevent the updated material composition information from becoming a negative value, and to perform correction so that the total value is 100% (1.0). Thus, according to the material development support system 100 according to the second embodiment, it is possible to explore appropriate material composition information while enjoying the same effects as the first embodiment.

[0318] [Third embodiment]

[0319] Although the default material composition information is not described in detail in the first and second embodiments, it can be assumed that the default material composition information cannot be used to search for material composition information having a phase fraction similar to the target phase fraction. Therefore, in the third embodiment, in order to search for material composition information having a phase fraction similar to the target phase fraction, the default material composition information is started from material composition information having a phase fraction relatively similar to the target phase fraction. The third embodiment will be described below with the differences from the first and second embodiments as the center.

[0320] <Functional structure of composition optimization device>

[0321] First, the functional structure of the composition optimization device of the third embodiment is described. Fig.24 FIG. 2 is another example of the functional structure of the composition optimization device. Fig.23 The functional structure of the composition optimization device 2300 described above is different in that the composition optimization device 2400 of the third embodiment has a selection unit 2401.

[0322] The selection unit 2401 acquires the input target phase fraction, reads material composition information similar to the acquired target phase fraction from the composition candidate storage unit 2402 , and inputs the material composition information into the material composition input unit 131 as default material composition information.

[0323] <Default material composition information>

[0324] Next, the default material composition information stored in the composition candidate storage unit 2402 will be described. Fig.25 This is a diagram showing an example of default material composition information. Fig.25 The example is an example of material composition information of aluminum alloy standards, and shows the maximum value and the minimum value of the allowable range of the ratio of each composition of 13 kinds of aluminum alloys.

[0325] In the selection unit 2401, one aluminum alloy is selected from the 13 aluminum alloys according to the set target phase fraction. In addition, in the selection unit 2401, the minimum value or maximum value of the allowable range of the ratio of each composition is set as the default material composition information for the selected one aluminum alloy.

[0326] <Results of composition optimization processing>

[0327] Next, differences in processing results of the composition optimization process caused by differences in default material composition information will be described. Fig.26 : is a diagram showing an example of default material composition information and a processing result of composition optimization processing. Fig. 27 FIG2 is a second diagram showing an example of a processing result of the composition optimization processing.

[0328] in, Fig.26 (a) shows the material composition information that achieves the target phase fraction (correct answer data = GT (GroundTruth (true value))), Fig. 27 (a) shows the target phase fraction.

[0329] also, Fig.26 (b) shows the processing result of the composition optimization processing when the selection unit 2401 selects a predetermined aluminum alloy as the default material composition information and takes the maximum value of the allowable range of the ratio of each composition as the default material composition information. Fig. 27 (b) shows the phase fraction at this time.

[0330] also, Fig.26 (c) shows the processing result of the composition optimization process when the selection unit 2401 selects a predetermined aluminum alloy as the default material composition information and the minimum value of the allowable range of the ratio of each composition is set as the default material composition information. Fig. 27 (c) shows the phase fraction at this time.

[0331] As from Fig. 27 (a) and Fig. 27 (b) and Fig. 27 (a) and Fig. 27 As is apparent from the comparison with (c), by selecting appropriate default material composition information in the selection unit 2401, a phase fraction similar to the target phase fraction can be obtained in any case. Fig.26 (a) and Fig.26 (b) and Fig.26 (a) and Fig.26 As is apparent from the comparison with (c), since the selection unit 2401 selects appropriate default material composition information, it is possible to search for material composition information similar to the correct answer data.

[0332] However, as from Fig.26 (a) and Fig.26 As is apparent from the comparison of (b) and (c), when the maximum value of the allowable range of the ratio of each component is used as the default material composition information, the composition ratio error of "Zn" is large. Fig.26 (a) and Fig.26 As is apparent from the comparison with (c), when the minimum value of the allowable range of the ratio of each component is used as the default material composition information, the error of the ratio of each component is small.

[0333] Thus, the reason why the composition ratio of "Zn" is different is that, in the case of aluminum alloy, "Zn" is in a solid solution state, does not precipitate phases, and does not appear in phase fractions, and is therefore affected by the default material composition information. That is, the ratio of the composition that does not appear in the phase fraction has low reliability and can be considered as the ratio that the user should ultimately adjust. Therefore, when the processing result of the composition optimization process is prompted to the user, the ratio of the composition that does not appear in the phase fraction can also be clearly indicated to the user as the ratio that the user should adjust.

[0334] Summary

[0335] As is clear from the above description, the composition optimization device of the third embodiment selects material composition information having a phase fraction similar to a target phase fraction as default material composition information and starts the composition optimization process.

[0336] Therefore, according to the composition optimization device of the third embodiment, it is possible to search for material composition information having a phase fraction similar to a target phase fraction.

[0337] It should be noted that in the above description, there is no mention of a user interface when selecting material composition information having a phase fraction similar to a target phase fraction. However, in the case of aluminum alloys, for example, for a plurality of standard alloys (aluminum alloys specified by standards) whose phase fractions are known, the phase fractions may be prompted to the user. Therefore, the user may select a phase fraction similar to the target phase fraction from the plurality of recommended phase fractions, and select the material composition information of the corresponding standard alloy as the default material composition information.

[0338] Alternatively, the user may be prompted with a standard alloy having a phase fraction similar to the target phase fraction as a candidate for the default material composition information. Alternatively, the user may be prompted with an allowable range of the ratio of each element of the standard alloy having a phase fraction similar to the target phase fraction. Therefore, the user may obtain the default material composition information by inputting the ratio of each element within the provided allowable range.

[0339] [Fourth embodiment]

[0340] In the third embodiment, the target phase fraction is set at each temperature in the predetermined temperature range. However, the method of setting the target phase fraction is not limited thereto, and the target phase fraction may be set only at a specific temperature (here, 100° C.) first.

[0341] In this case, the selection unit 2401 determines the material composition information that achieves the target phase fraction at 100° C., and sets the determined material composition information as the default material composition information. In this case, the simulation unit 503 operates the simulator for the set default material composition information, and calculates the phase fraction at each temperature within the predetermined temperature range (except the phase fraction at 100° C.).

[0342] Then, the phase fractions at each temperature (except 100° C.) within the predetermined temperature range calculated by the simulation unit 503 and the initially set target phase fraction at 100° C. are set as the target phase fractions at each temperature within the predetermined temperature range.

[0343] Therefore, the composition optimization device 2400 can explore the material composition information that achieves the target phase fraction at each temperature within the set predetermined temperature range based on the set default material composition information.

[0344] [Fifth embodiment]

[0345] In each of the above-mentioned embodiments, the phase fraction error (logarithmic value) calculation unit 903 compares the phase fraction (correct data) at each temperature of 100°C to 700°C with the phase fraction (output data) at each temperature of 100°C to 700°C, and calculates the error of the logarithmic value of the phase fraction at each temperature. However, the method of calculating the logarithmic value error by the phase fraction error (logarithmic value) calculation unit 903 is not limited to this. For example, when calculating the logarithmic value of the phase fraction, the logarithmic value of the phase fraction may be non-negative by adding the value corresponding to the first decimal place of the phase fraction, and then calculating the error.

[0346] Furthermore, in the above-mentioned embodiments, although the prediction direction when the prediction device 120 "predicts the phase fraction sequentially at predetermined temperature intervals" is not particularly described, the prediction direction may be either a temperature increase direction or a temperature decrease direction. Alternatively, it may be both a temperature increase direction and a temperature decrease direction.

[0347] In addition, in each of the above embodiments, as a specific example of the material composition, the alloy composition representing the ratio of the metal element contained in the alloy to the added element (other metal element or non-metal element) is described, but the material composition is not limited to the alloy composition. For example, it may be a chemical composition representing the ratio of each chemical component contained in a substance other than the alloy.

[0348] In addition, in each of the above embodiments, as a trained prediction model, the case of using either Seq2Seq or Transformer with an Attention mechanism is described. However, the trained prediction model is not limited to this, and other architectures can be used as long as they are architectures that can calculate time series data as data at predetermined time intervals. Specifically, RNN (Recurrent Neural Network), Bidirectional RNN, Seq2Seq, etc. can also be applied. Alternatively, GRU (Gated Recurrent Unit), LSTM (Long Short Term Memory), etc. can also be applied.

[0349] In addition, in each of the above embodiments, the training device 110 and the prediction device 120 are configured as separate bodies, but the training device 110 and the prediction device 120 may be configured as an integral body. Similarly, in each of the above embodiments, the training device 110 and the composition optimization device 130 are configured as separate bodies, but the training device 110 and the composition optimization device 130 may be configured as an integral body. Similarly, in each of the above embodiments, the prediction device 120 and the composition optimization device 130 are configured as separate bodies, but the prediction device 120 and the composition optimization device 130 may be configured as an integral body.

[0350] Furthermore, although the application scenarios of the exploration results of the material composition information according to the target phase fraction are not mentioned in the above embodiments, for example, material design and material development can be performed based on the exploration results of the material composition information according to the target phase fraction.

[0351] It should be noted that the present invention is not limited to the structures illustrated in the above embodiments, and the structures shown here such as combinations thereof with other elements, etc. These points can be changed within the scope of the present invention and can be appropriately predetermined according to the application form.

[0352] This application claims priority based on Japanese invention patent application No. 2022-166274 filed on October 17, 2022, and all the contents of the Japanese patent application are incorporated herein by reference.

[0353] Description of Reference Numerals

[0354] 100: Material development support system

[0355] 110: Training device

[0356] 111: Training data generation unit

[0357] 112: Training Unit

[0358] 120: Prediction device

[0359] 121: Material composition input unit

[0360] 122: Prediction unit

[0361] 123: Display unit

[0362] 130: Composition Optimization Device

[0363] 131: Material composition input unit

[0364] 132: Prediction unit

[0365] 133: Loss function calculation unit

[0366] 134: Update unit

[0367] 400: Training data

[0368] 501: Element information input unit

[0369] 502: Combination decision unit

[0370] 503: Analog unit

[0371] 504: Save control unit

[0372] 601: Encoder unit

[0373] 602: Decoder unit

[0374] 603: Loss function calculation unit

[0375] 901: Phase fraction error calculation unit at generation / disappearance temperature

[0376] 902: Phase ratio error calculation unit

[0377] 903: Phase ratio error (logarithmic value) calculation unit

[0378] 904: Phase ratio error (difference value) calculation unit

[0379] 905: Cross entropy error calculation unit

[0380] 906: Weighted addition unit

[0381] 1101: Trained encoder unit

[0382] 1102: Encoder unit trained

[0383] 2300: Composition Optimization Device

[0384] 2301: Constraint Unit

[0385] 2400: Composition Optimization Device

[0386] 2401: Select Unit

Claims

1. A composition optimization device, include: a prediction unit that predicts the phase fraction of a material having the predetermined material composition at each temperature within a predetermined temperature range by inputting a predetermined material composition into a trained model trained using training data, wherein the material composition of a training target material and the phase fraction of the training target material at each temperature within the predetermined temperature range are associated with each other in the training data; as well as an updating unit that updates the predetermined material composition input into the trained model by back-propagating an error calculated based on the target phase fraction and the predicted phase fraction, The process of predicting the phase fraction based on the updated material composition by the prediction unit and the process of updating the predetermined material composition by back-propagating the calculated error by the update unit are repeatedly performed until the calculated error satisfies a predetermined condition.

2. The composition optimization device according to claim 1, in, The trained model is configured to predict the phase fraction at the (i+1)th temperature using the phase fractions up to the i-th temperature within the predetermined temperature range predicted by the trained model, where i is an integer greater than or equal to 1.

3. The composition optimization device according to claim 2, in, The trained model adopts a structure capable of calculating data at predetermined time intervals, namely, time series data. The trained model predicts the phase fraction at predetermined temperature intervals based on the predetermined material composition of the material.

4. The composition optimization device according to claim 3, in, The trained model is any one of RNN, bidirectional RNN, Seq2Seq, Seq2Seq with Attention mechanism, GRU, LSTM and Transformer.

5. The composition optimization device according to claim 4, in, The trained model includes: an encoder unit that outputs a feature amount by inputting a material composition of a predetermined material; and A decoder unit predicts the phase fraction at the (i+1)th temperature by inputting the output feature quantity and the predicted phase fractions up to the i-th temperature.

6. The composition optimization device according to any one of claims 1 to 5, in, The phase fraction is the phase fraction in a thermodynamic equilibrium state.

7. The composition optimization device according to any one of claims 1 to 6, in, The invention further includes: a constraint unit that corrects the updated predetermined material composition in such a manner that the updated predetermined material composition updated by the updating unit satisfies the constraint related to the material composition.

8. The composition optimization device according to claim 7, in, The constraint unit corrects the updated predetermined material composition so that the total value of the material composition is 100% and the ratio of each composition is 0% or more.

9. The composition optimization device according to any one of claims 1 to 8, in, The system further includes a loss function calculation unit that calculates an error calculated based on the target phase fraction and the predicted phase fraction.

10. The composition optimization device according to claim 9, in, The error calculated by the loss function calculation unit includes at least any one of the following: a first addition result obtained by adding the predicted phase fraction at the temperature at which each phase is generated or disappeared determined based on the target phase fraction for all phases and an error between the target phase fraction; a second addition result obtained by adding the error between the predicted phase fraction at each temperature within a predetermined temperature range and the target phase fraction at each temperature; a third addition result obtained by adding the error between the predicted logarithmic value of the phase fraction at each temperature within a predetermined temperature range and the target logarithmic value of the phase fraction at each temperature; a fourth addition result obtained by adding a difference value of the phase fraction between adjacent temperatures in the predicted phase fraction at each temperature within a predetermined temperature range and an error between the difference value of the phase fraction between adjacent temperatures in the predicted phase fraction at each temperature; as well as The fifth addition result is obtained by adding the error between the predicted ratio of the phase fractions at each temperature within a predetermined temperature range and the target ratio of the phase fractions at each temperature.

11. The composition optimization device according to claim 10, in, The loss function calculation unit performs weighted addition on the first to fifth addition results.

12. The composition optimization device according to claim 10, in, When calculating the logarithmic value of the phase fraction, the loss function calculation unit makes the logarithmic value of the phase fraction a non-negative value by adding a value corresponding to the first decimal place of the phase fraction.

13. The composition optimization device according to any one of claims 1 to 12, in, The predetermined material composition is a material composition selected from material compositions specified by a standard and having a phase fraction similar to a target phase fraction.

14. A composition optimization method, in, The following steps are performed by a computer constituting the optimization device: A prediction step of predicting a phase fraction of a material having the predetermined material composition at each temperature within a predetermined temperature range by inputting a predetermined material composition into a trained model trained using training data, wherein the material composition of a training target material and the phase fraction of the training target material at each temperature within the predetermined temperature range are associated with each other in the training data; as well as an updating step of updating the predetermined material composition input into the trained model by back-propagating an error calculated based on the target phase fraction and the predicted phase fraction, The process of predicting the phase fraction based on the updated material composition in the prediction step and the process of updating the predetermined material composition by back-propagating the calculated error in the update step are repeatedly performed until the calculated error satisfies a predetermined condition.

15. A composition optimization program, which causes a computer of a composition optimization device to execute the following steps: a prediction step of predicting the phase fraction of a material having the predetermined material composition at each temperature within a predetermined temperature range by inputting a predetermined material composition into a trained model trained using training data in which the material composition of a training target material and the phase fraction of the training target material at each temperature within the predetermined temperature range are associated; and an updating step of updating the predetermined material composition input into the trained model by back-propagating an error calculated based on the target phase fraction and the predicted phase fraction, The process of predicting the phase fraction based on the updated material composition in the prediction step and the process of updating the predetermined material composition by back-propagating the calculated error in the update step are repeatedly performed until the calculated error satisfies a predetermined condition.

Images