Prediction result analysis device, analysis method, design method, and prediction program

Abstract

To facilitate grasp of a degree of influence on a phase fraction associated with changing a material composition.SOLUTION: A prediction result analysis device has: a storage unit that stores a learned model trained by using learning data in which a material composition and a phase fraction of a material to be learned are associated with each other; an intermediate composition generation unit that generates the material composition of an intermediate material; a prediction unit that, in predicting a phase fraction by inputting the material composition of the intermediate material to the learned model, partially differentiates, at each temperature, a value predicted for a predetermined phase by using the material composition of the intermediate material to calculate a partial differential value; an integral gradient calculation unit that integrates, at each temperature, the calculated partial differential value for each component included in the material composition of the intermediate material to calculate an integral gradient of each temperature and each component; and a display unit that, on a heat map with the components arranged on a first axis and the temperatures arranged on a second axis, displays areas identified on the basis of the components and the temperatures in a display mode according to the integral gradients.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a prediction result analysis device, an analysis method, and an analysis program. [Background technology]

[0002] In designing materials, it is important to calculate the phase fraction in a thermodynamic equilibrium state over a specific temperature range, and surrogate models that predict the phase fraction based on material composition information are being developed. As an example, a surrogate model that predicts the phase fraction based on the content of added elements in an alloy is being developed.

[0003] Here, the phase fraction in a specific temperature range correlates with the alloy properties, so when a designer designs an alloy using the above model, it is important for the designer to know in advance which phase in which temperature range can be affected by increasing or decreasing the content of which added element to approach the desired phase fraction, in order to design an alloy efficiently. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] International Publication No. 2020 / 090617 Summary of the Invention [Problem to be solved by the invention]

[0005] In response to this, for example, a designer may predict the phase fraction within a specific temperature range while sequentially increasing or decreasing the content of all additive elements, and then compare the predicted phase fraction with the phase fraction before the change. By performing such a procedure, the designer can grasp in advance the degree of effect of the content of each additive element on the phase fraction.

[0006] However, such an operation places a heavy burden on the designer when the alloy contains many additive elements, and there is also a problem that it is difficult for the designer to grasp the degree of influence of the content of each additive element on the phase fraction by simply comparing the phase fractions.

[0007] The present disclosure aims to make it easier for designers to understand the degree of effect on phase fractions associated with changes in material composition. [Means for solving the problem]

[0008] A prediction result analysis device according to a first aspect of the present disclosure, a memory unit that stores a trained model that is trained using training data in which a material composition of a training target material is associated with a phase fraction of the training target material at each temperature within a specific temperature interval, the trained model predicting a phase fraction at an (i+1)th temperature using phase fractions predicted by the trained model for each temperature up to an i-th temperature (i is an integer equal to or greater than 1) within the specific temperature interval; an intermediate composition generating unit that divides a material composition of a material to be predicted and a material composition of a baseline material to generate material compositions of a plurality of intermediate materials; a prediction unit that, when predicting a phase fraction at each temperature of each of the plurality of intermediate materials in the specific temperature range by inputting material compositions of the plurality of intermediate materials into the trained model, calculates a partial differential value by partially differentiating the predicted value for a pre-specified phase at each temperature using each of the material compositions of the plurality of intermediate materials; an integral gradient calculation unit that calculates an integral gradient for each temperature and each component by integrating the calculated partial differential value for each component included in the material composition of the intermediate materials at each temperature; The heat map has each component arranged on a first axis and each temperature arranged on a second axis, and has a display unit that displays each area identified based on each component and each temperature in a display mode corresponding to the integral gradient.

[0009] A second aspect of the present disclosure is a prediction result analysis device according to the first aspect, The integral gradient calculation unit calculates the integral gradient for each temperature and each component by integrating the calculated partial differential values ​​for each component contained in the material composition of the multiple intermediate materials at each temperature and then averaging the results.

[0010] A third aspect of the present disclosure is a prediction result analysis device according to the second aspect, When integrating the calculated partial differential value for each component contained in the material composition of the plurality of intermediate materials at each temperature, the integral gradient calculation unit integrates the partial differential value using a composition weight calculated in advance for each of the plurality of intermediate materials.

[0011] A fourth aspect of the present disclosure is a prediction result analysis device according to the third aspect, The integral gradient calculation unit calculates the composition weight based on a difference between each of the material compositions of the plurality of intermediate materials and a material composition of the prediction target material.

[0012] A fifth aspect of the present disclosure is a prediction result analysis device according to any one of the first to fourth aspects, The intermediate composition generating unit generates material compositions of the intermediate materials for each of a plurality of baseline material compositions, The integral gradient calculation unit calculates the integral gradient of each temperature and each component for each material composition of the plurality of baseline materials; The display unit displays the heat map in a display mode according to an averaged integral gradient obtained by averaging a plurality of the integral gradients for each temperature and each component.

[0013] A sixth aspect of the present disclosure is a prediction result analysis device according to any one of the first to fifth aspects, The trained model is applied with an architecture capable of processing time series data, which is data for each specific time interval, and the trained model predicts the phase fraction for each specific temperature interval based on the material composition of the material to be predicted.

[0014] A seventh aspect of the present disclosure is a prediction result analysis device according to the sixth aspect, The trained model is any one of RNN, Bidirectional RNN, Seq2Seq, Seq2Seq with Attention mechanism, GRU, LSTM, and Transformer.

[0015] An eighth aspect of the present disclosure is a prediction result analysis device according to the seventh aspect, The trained model is an encoder unit that outputs a feature quantity in response to an input of a material composition of a material to be predicted; The decoder unit predicts the phase fraction at the (i+1)th temperature by inputting the output feature amount and the predicted phase fractions at the temperatures up to the i-th temperature.

[0016] A ninth aspect of the present disclosure is a prediction result analysis device according to any one of the first to eighth aspects, The phase fraction is a phase fraction in a thermodynamic equilibrium state.

[0017] An analysis method according to a tenth aspect of the present disclosure includes: An analysis method in which a trained model is trained using training data in which a material composition of a training target material and a phase fraction of the training target material at each temperature within a specific temperature interval are associated with each other, the trained model predicting a phase fraction at an (i+1)th temperature using phase fractions predicted by the trained model for temperatures up to an i-th temperature (i is an integer equal to or greater than 1) within the specific temperature interval, the method comprising: The computer, A step of dividing the material composition of the material to be predicted and the material composition of the baseline material to generate a plurality of intermediate material compositions; a step of inputting material compositions of the intermediate materials into the trained model to predict phase fractions at each temperature of the intermediate materials in the specific temperature range, partially differentiating the predicted values ​​for a pre-specified phase at each temperature using the material compositions of the intermediate materials to calculate partial differential values; calculating an integral gradient for each temperature and each component by integrating the calculated partial differential value for each component contained in the material composition of the intermediate materials at each temperature; In a heat map in which each component is arranged on a first axis and each temperature is arranged on a second axis, a process is carried out in which each area identified based on each component and each temperature is displayed in a display manner corresponding to the integral gradient.

[0018] A prediction program according to an eleventh aspect of the present disclosure, A trained model is trained using training data in which a material composition of a training target material is associated with a phase fraction at each temperature of the training target material within a specific temperature interval, and the trained model predicts a phase fraction at an (i+1)th temperature using phase fractions predicted by the trained model for each temperature up to an i-th temperature (i is an integer equal to or greater than 1) within the specific temperature interval, said trained model being a prediction program for causing a computer to execute the trained model, The computer includes: A step of dividing the material composition of the material to be predicted and the material composition of the baseline material to generate a plurality of intermediate material compositions; a step of inputting material compositions of the intermediate materials into the trained model to predict phase fractions at each temperature of the intermediate materials in the specific temperature range, partially differentiating the predicted values ​​for a pre-specified phase at each temperature using the material compositions of the intermediate materials to calculate partial differential values; calculating an integral gradient for each temperature and each component by integrating the calculated partial differential value for each component contained in the material composition of the intermediate materials at each temperature; In a heat map in which each component is arranged on a first axis and each temperature is arranged on a second axis, a process is executed in which each area identified based on each component and each temperature is displayed in a display manner corresponding to the integral gradient. Effect of the Invention

[0019] According to the present disclosure, designers can easily grasp the degree of effect on phase fractions associated with changes in material composition. [Brief description of the drawings]

[0020] [Figure 1] FIG. 1 is a diagram illustrating an example of a system configuration of a prediction system and an example of functional configurations of a learning device, a prediction device, and a prediction result analysis device. [Diagram 2] FIG. 2 is a diagram illustrating an example of a hardware configuration of the learning device, the prediction device, and the prediction result analysis device. [Diagram 3] FIG. 3 is a diagram showing a configuration of learning data and specific examples of input data and correct answer data. [Figure 4] FIG. 4 is a diagram illustrating an example of a functional configuration of the learning data generating unit. [Diagram 5] FIG. 5 is a diagram illustrating an example of a functional configuration of the learning unit. [Figure 6] FIG. 6 is a first diagram illustrating an example of the operation of the learning unit. [Figure 7] FIG. 7 is a second diagram illustrating an example of the operation of the learning unit. [Figure 8] FIG. 8 is a diagram illustrating an example of a method for calculating an integrated loss by the loss function calculation unit. [Figure 9] FIG. 9 is a flowchart showing the flow of the learning process. [Figure 10] FIG. 10 is a first diagram illustrating an example of a functional configuration of the prediction unit. [Figure 11] FIG. 11 is a first diagram illustrating an example of the operation of the prediction unit. [Figure 12] FIG. 12 is a flowchart showing the flow of the prediction process. [Figure 13]FIG. 13 is a diagram for explaining the prediction accuracy of the phase fraction. [Figure 14] FIG. 14 is a first diagram showing a specific example of the process of the prediction result analysis device. [Figure 15] FIG. 15 is a second diagram illustrating an example of the functional configuration of the prediction unit. [Figure 16A] FIG. 16A is a second diagram illustrating an example of the operation of the prediction unit. [Figure 16B] FIG. 16B is a third diagram illustrating an example of the operation of the prediction unit. [Figure 16C] FIG. 16C is a fourth diagram illustrating an example of the operation of the prediction unit. [Figure 17] FIG. 17 is a second diagram showing a specific example of the process of the prediction result analysis device. [Figure 18] FIG. 18 is a flowchart showing the flow of the prediction result analysis process. [Figure 19] FIG. 19 is a flowchart showing the flow of the baseline composition information input process. [Figure 20] FIG. 20 is a flowchart showing the flow of the integral gradient calculation process. [Figure 21] FIG. 21 is a diagram showing an example of the material composition of the prediction target material used in the prediction result analysis process and the predicted phase fraction. [Figure 22] FIG. 22 is a diagram showing an example of a heat map displayed in a display mode according to the integral gradient. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] Hereinafter, each embodiment will be described with reference to the accompanying drawings. In this specification and the drawings, components having substantially the same functional configurations are denoted by the same reference numerals, and redundant description will be omitted.

[0022] [First embodiment] <System configuration of prediction system, functional configuration of learning device, prediction device, and prediction result analysis device> First, a system configuration of a prediction system including a prediction result analysis device according to a first embodiment, and functional configurations of a learning device, a prediction device, and a prediction result analysis device included in the prediction system will be described. In this embodiment, the prediction system is a system that predicts phase fractions at each temperature within a specific temperature interval based on material composition information. The phase fractions referred to here are phase fractions in a thermodynamic equilibrium state, and hereinafter, the phase fractions in a thermodynamic equilibrium state will be simply referred to as "phase fractions."

[0023] 1 is a diagram showing an example of a system configuration of a prediction system and an example of a functional configuration of a learning device, a prediction device, and a prediction result analysis device. As shown in FIG. 1, the prediction system 100 includes a learning device 110, a prediction device 120, and a prediction result analysis device 130.

[0024] A learning program is installed in the learning device 110, and the learning device 110 functions as a learning data generating unit 111 and a learning unit 112 by executing the program.

[0025] The learning data generation unit 111 generates learning data for learning a prediction model, and then stores the generated learning data in the learning data storage unit 113. Note that the prediction model referred to here is a surrogate model that replaces a conventional model with machine learning. In this embodiment, the learning data storage unit 113 stores the following as learning data for learning a prediction model: -Multiple material composition information with different combinations of content rates of each component The phase fraction at each temperature within a specific temperature interval for each material having material composition information; are stored in association with each other.

[0026] The learning unit 112 reads out the learning data from the learning data storage unit 113, and then uses the read learning data to learn the prediction model. The learning unit 112 learns the prediction model in order to bring the output data output by inputting the material composition information stored in the learning data into the prediction model closer to the phase fraction at each temperature within a specific temperature interval that is stored in association with the learning data. In this way, the learning unit 112 generates a learned prediction model. The learning unit 112 stores the generated learned prediction model in the storage unit of the prediction unit 122 of the prediction device 120 and the storage unit of the prediction unit 134 of the prediction result analysis device 130.

[0027] In this embodiment, the learning unit 112 has an architecture capable of processing time-series data, which is data for each specific time interval, as a prediction model, and is generally used in fields such as natural language processing. Seq2Seq with attention mechanism, or ·Transformer, In this way, according to the prediction model, by inputting the material composition information stored in the learning data, it is possible to sequentially output output data corresponding to the phase fraction at each temperature within a specific temperature interval.

[0028] Here, "sequentially outputting output data corresponding to the phase fraction at each temperature" means that the prediction model outputs output data corresponding to the phase fraction at the (i+1)th temperature by using the correct data at each temperature up to the i-th temperature, where i is an integer of 1 or more.

[0029] In this way, the learning unit 112 is configured to sequentially output output data corresponding to the phase fraction at each temperature. In other words, the learning unit 112 is configured to process the output data corresponding to the phase fraction at each temperature as time-series data. This allows the learning unit 112 to perform learning that reflects the output data corresponding to the phase fraction between adjacent temperature regions.

[0030] A prediction program is installed in the prediction device 120, and the prediction device 120 functions as a material composition input unit 121, a prediction unit 122, and a display unit 123 by executing the program.

[0031] The material composition input unit 121 accepts input of material composition information of the material to be predicted, and notifies the prediction unit 122 of the information.

[0032] The prediction unit 122 stores in the storage unit the trained prediction model trained by the learning unit 112. After reading out the trained prediction model from the storage unit, the prediction unit 122 inputs the material composition information notified from the material composition input unit 121 to the trained prediction model, thereby sequentially predicting the phase fraction at each temperature within a specific temperature interval. In other words, the prediction unit 122 sequentially predicts the phase fraction within a specific temperature interval for each specific temperature interval.

[0033] Here, "sequentially predicting the phase fraction at specific temperature intervals" refers to, for example, a trained prediction model predicting the phase fraction at the (i+1)th temperature using the phase fractions predicted by the trained prediction model for each temperature up to the i-th temperature (i is an integer greater than or equal to 1).

[0034] Thus, in this embodiment, the prediction unit 122 has a configuration for sequentially predicting the phase fraction at each temperature. That is, the prediction unit 122 has a configuration for processing the phase fraction at each temperature as time-series data by a recurrent network. As a result, the prediction unit 122 can perform prediction that reflects the predicted data of the phase fraction between adjacent temperature regions. As a result, it is possible to suppress a decrease in the prediction accuracy of the phase fraction compared to, for example, a case where a multilayer neural network is applied, that is, a case where a network in which the predicted result of the phase fraction between adjacent temperature regions is not reflected is applied. In other words, the prediction unit 122 can improve the prediction accuracy when predicting the phase fraction over a specific temperature range based on the material composition.

[0035] The display unit 123 displays the phase fraction at each temperature within a specific temperature interval predicted by the prediction unit 122. The display unit 123 displays the phase fraction at each temperature within a specific temperature interval predicted by the prediction unit 122 in a different color for each phase. However, the phase fraction at each temperature within a specific temperature interval predicted by the prediction unit 122 may be displayed using different line types for each phase, such as a solid line, a dotted line, or a dashed dotted line.

[0036] An analysis program is installed in the prediction result analysis device 130. By executing the program, the prediction result analysis device 130 functions as a material composition input unit 131, a baseline composition input unit 132, an intermediate composition generation unit 133, a prediction unit 134, an integral gradient calculation unit 135, and a display unit 136.

[0037] The material composition input unit 131 accepts input of material composition information of the material to be predicted, and notifies the intermediate composition generating unit 133. Note that the material composition information of the material to be predicted input to the material composition input unit 131 is the same as the material composition information of the material to be predicted input to the material composition input unit 121 of the prediction device 120.

[0038] The baseline composition input unit 132 accepts input of baseline composition information, which is material composition information of a material that becomes a baseline, and notifies the intermediate composition generation unit 133. The baseline composition information is material composition information used to calculate gradients when visualizing the basis of prediction using the Integrated Gradients method.

[0039] For example, it is preferable to use representative material composition information of the material to be predicted, or material composition information including the upper limit or lower limit of the composition of the material to be predicted, as the baseline composition information. Specifically, in the case of 6000 series aluminum alloys, there are JIS standards such as A6061 and A6063, and the upper limit and lower limit of the composition of the additive elements are determined for each standard. In this case, it is preferable to use material composition information including the upper limit or lower limit of the additive elements in the standard to which the material to be predicted belongs as the baseline composition information.

[0040] Furthermore, when a particular precipitation phase is of interest, material composition information in which it is known that the particular precipitation phase occurs may be used as baseline composition information.

[0041] By using multiple baseline composition information, it is possible to appropriately evaluate the degree of influence on the phase fraction even at a phase fraction where a specific additive element is saturated.

[0042] The intermediate composition generating unit 133 generates intermediate composition information, which is material composition information of the intermediate material, based on the material composition information of the material to be predicted notified from the material composition input unit 131 and the baseline composition information notified from the baseline composition input unit 132. Specifically, the intermediate composition generating unit 133 generates the intermediate composition information by dividing the gap between the material composition information of the material to be predicted and the baseline composition information by, for example, the number of steps=M. As a result, the intermediate composition generating unit 133 can generate M pieces of intermediate composition information that gradually approach the baseline composition information from the material composition information of the material to be predicted.

[0043] The intermediate composition generating unit 133 sequentially inputs the material composition information of the multiple intermediate materials thus generated, that is, M pieces of intermediate composition information, to the predicting unit 134.

[0044] The prediction unit 134 stores in the storage unit the trained prediction model trained by the learning unit 112. After reading out the trained prediction model from the storage unit, the prediction unit 134 inputs the M pieces of intermediate composition information notified by the intermediate composition generating unit 133 to the trained prediction model. As a result, the prediction unit 134 sequentially predicts the phase fraction at each temperature within a specific temperature interval. In other words, the prediction unit 122 sequentially predicts the phase fraction within a specific temperature interval for the M pieces of intermediate composition information for each specific temperature interval.

[0045] When predicting the phase fraction within a predetermined temperature range for each predetermined temperature interval, the prediction unit 134 partially differentiates the predicted value for the pre-specified phase at each temperature using each of the M pieces of intermediate composition information. In this way, the prediction unit 134 can calculate a "partial differential value" that indicates the degree to which the pre-specified phase changes at each temperature by changing each of the M pieces of intermediate composition information by a small amount.

[0046] The integral gradient calculation unit 135 calculates the integral gradient for each temperature and each component by integrating the partial differential value of each of the M pieces of intermediate composition information for each temperature and each component.

[0047] In a heat map in which each component is arranged on the horizontal axis and each temperature is arranged on the vertical axis, the display unit 136 arranges each region specified based on each component and each temperature in a display manner according to the integral gradient. Note that the horizontal axis is an example of a first axis, and the vertical axis is an example of a second axis. The display unit 136 displays a heat map in which colors are arranged in a display manner according to the integral gradient.

[0048] This allows a material designer to easily understand, based on the heat map displayed on the prediction result analysis device 130, how to change the material composition to approach the desired phase fraction relative to the phase fraction displayed on the prediction device 120.

[0049] In other words, according to the prediction result analysis device 130, when a material designer predicts a phase fraction based on material composition information, the designer can easily grasp, based on a heat map, the degree of effect on the phase fraction due to a change in the material composition.

[0050] <Hardware configuration of the learning device, prediction device, and prediction result analysis device> Next, a description will be given of the hardware configurations of the learning device 110, the prediction device 120, and the prediction result analysis device 130. Since the learning device 110, the prediction device 120, and the prediction result analysis device 130 have similar hardware configurations, the hardware configurations of the learning device 110, the prediction device 120, and the prediction result analysis device 130 will be collectively described here with reference to FIG.

[0051] Fig. 2 is a diagram showing an example of the hardware configuration of the learning device, the prediction device, and the prediction result analysis device. As shown in Fig. 2, the learning device 110, the prediction device 120, and the prediction result analysis device 130 each include a processor 201, a memory 202, an auxiliary storage device 203, an interface device 204, a communication device 205, and a drive device 206. The hardware components of the learning device 110, the prediction device 120, and the prediction result analysis device 130 are connected to each other via a bus 207.

[0052] The processor 201 has various arithmetic devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 201 reads various programs onto the memory 202 and executes them. Note that the various programs referred to here include, for example, a learning program, a prediction program, an analysis program, and other programs.

[0053] The memory 202 has 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. The computer realizes the above-mentioned various functions by the processor 201 executing various programs read onto the memory 202.

[0054] The auxiliary storage device 203 stores various programs and various data used when the various programs are executed by the processor 201. For example, the learning data storage unit 113 is realized in the auxiliary storage device 203. Alternatively, a storage unit that stores a trained prediction model is realized in the auxiliary storage device 203.

[0055] The interface unit 204 is a connection device that connects to an operation unit 211, which is an example of a user interface unit, and a display unit 212. The communication unit 205 is a communication device for communicating with an external device (not shown) via a network.

[0056] The drive device 206 is a device for setting the recording medium 213. The recording medium 213 here includes media that optically, electrically, or magnetically record information, such as a CD-ROM, a flexible disk, and a magneto-optical disk. The recording medium 213 may also include media such as semiconductor memory that electrically records information, such as a ROM or a flash memory.

[0057] The various programs to be installed in the auxiliary storage device 203 are installed, for example, by setting the distributed recording medium 213 in the drive device 206 and reading out the various programs recorded in the recording medium 213 by the drive device 206. Alternatively, the various programs to be installed in the auxiliary storage device 203 may be installed by being downloaded from a network via the communication device 205.

[0058] <Learning data configuration, input data, and correct answer data> Next, a configuration of learning data and specific examples of input data and correct answer data included in the learning data will be described. Fig. 3 is a diagram showing the configuration of learning data and specific examples of input data and correct answer data.

[0059] As shown in FIG. 3, the learning data 300 includes, as information items, "input data" and "correct answer data."

[0060] The "input data" stores "material composition information 1", "material composition information 2", ..., which are material composition information with different combinations of content rates of each component. In Fig. 3, reference numeral 301 indicates a specific example of "material composition information 1", which shows an example of the content rates of each component constituting "6013", which is the name of a known aluminum alloy standard, specifically, each of the added elements Si, Fe, Cu, Mn, Mg, Cr, Zn, and Ti.

[0061] Similarly, reference numeral 302 indicates a specific example of "material composition information 2," which indicates the content of each component constituting "6060," a known aluminum alloy standard designation, specifically, the content of each of the added elements Si, Fe, Cu, Mn, Mg, Cr, Zn, and Ti.

[0062] The "correct data" stores "phase fraction 1", "phase fraction 2", ..., which are phase fractions at each temperature in a specific temperature range, associated with "material composition information 1", "material composition information 2", ..., etc. In the example of Fig. 3, the specific temperature range is from 100°C to 800°C, and both 100°C and 800°C are included in the specific temperature range.

[0063] In FIG. 3, reference numeral 311 indicates a specific example of "phase fraction 1," and reference numeral 312 indicates a specific example of "phase fraction 2." In both examples, the horizontal axis indicates temperature, and the vertical axis indicates phase fraction.

[0064] Here, reference numerals 311 and 312 indicate results obtained by operating a high-cost simulator with high prediction accuracy for a long period of time, and in this embodiment, these results are used as correct answer data.

[0065] The simulator here refers to software that calculates the phase fraction at each temperature within a specific temperature range by thermodynamic equilibrium calculation. Specifically, the simulator includes software such as CaTCalc and MatCalc. Alternatively, the simulator may include software such as Termosuite, FactStage, and Pandat. Alternatively, the simulator may include software such as MALT2, Thermo-Calc, and OpenCalphad.

[0066] <Details of each part of the learning device> Next, the learning data generating unit 111 and the learning unit 112, which are the components of the learning device 110, will be described in detail.

[0067] (1) Functional configuration of the learning data generation unit First, a description will be given of the functional configuration of the learning data generation unit 111. Fig. 4 is a diagram showing an example of the functional configuration of the learning data generation unit.

[0068] As shown in FIG. 4, the learning data generation unit 111 includes an element information input unit 401 , a combination determination unit 402 , a simulation unit 403 , and a storage control unit 404 .

[0069] The element information input unit 401 receives input of, for example, the type of additive element added when generating a specific alloy as each component included in the material composition information. The element information input unit 401 receives input of the upper and lower limits of the content rate for each additive element.

[0070] The combination determination unit 402 randomly selects the content rate of each added element under the constraints of the upper and lower limits for each added element that has been input, thereby determining multiple combinations of the content rates of each added element. The combinations of the content rates of each added element may be created by dividing the content rate between the upper and lower limits by a specific interval. The combination determination unit 402 notifies the simulation unit 403 of material composition information indicated by the multiple combinations that have been determined.

[0071] The simulation unit 403 calculates the phase fraction at each temperature in a specific temperature range by operating the above-mentioned simulator for each of the multiple pieces of material composition information notified from the combination determination unit 402. The simulation unit 403 notifies the storage control unit 404 of the multiple pieces of material composition information and the corresponding phase fraction at each temperature in the specific temperature range.

[0072] The storage control unit 404 generates learning data in which the multiple pieces of material composition information notified from the simulation unit 403 are used as "input data" and the phase fractions at each temperature within a corresponding specific temperature interval are used as "correct answer data." The storage control unit 404 stores the generated learning data in the learning data storage unit 113.

[0073] (2) Functional configuration of the learning unit Next, a description will be given of the functional configuration of the learning unit 112. Fig. 5 is a diagram showing an example of the functional configuration of the learning unit.

[0074] As shown in Fig. 5, the learning unit 112 has a prediction model and a loss function calculation unit 503. The prediction model is a surrogate model that replaces a conventional model with machine learning. As described above, the prediction model applies Seq2Seq or Transformer with an Attention mechanism, and includes an encoder unit 501 and a decoder unit 502.

[0075] The encoder unit 501 receives “material composition information 1”, “material composition information 2”, . . . stored in the “input data” of the learning data 300, and outputs feature amounts.

[0076] The decoder unit 502 receives the feature values ​​output by the encoder unit 501 and sequentially outputs output data. Specifically, the decoder unit 502 reads out "phase fraction 1," "phase fraction 2," ... etc. stored in the "correct answer data" of the learning data 300. Next, the decoder unit 502 uses the phase fractions at each temperature up to the i-th temperature, that is, the correct answer data at each temperature up to the i-th temperature, to output output data at the (i+1)-th temperature. In this way, the decoder unit 502 can, for example, The feature quantity output from the encoder unit 501 in response to the input of "material composition information 1"; "Correct data" for "phase fraction 1" and Based on the above, the decoder unit 502 can sequentially output output data corresponding to the phase fraction at each temperature in a specific temperature range. The feature quantity output from the encoder unit 501 in response to the input of "material composition information 2"; "Correct data" for "phase fraction 2" and Based on this, output data corresponding to the phase fraction at each temperature within a specific temperature interval can be sequentially output.

[0077] It is assumed that a softmax function is used in the output layer of the decoder unit 502. As a result, the multiple values ​​included in the output data of the decoder unit 502, i.e., the values ​​indicating the ratios of each phase, are in the range of 0 to 1, and the sum of the multiple values, i.e., the sum of the values ​​indicating the ratios of each phase, is "1".

[0078] In this way, by using a softmax function in the output layer of the decoder unit 502, output data corresponding to the phase fraction at each temperature within a specific temperature interval can be output sequentially without performing additional calculations.

[0079] The loss function calculation unit 503 reads out “phase fraction 1”, “phase fraction 2”, …, etc. stored in the “correct data” of the learning data 300. The loss function calculation unit 503 compares the read out phase fraction 1”, “phase fraction 2”, …, etc. with the output data at the (i+1)th temperature output from the decoder unit 502.

[0080] In comparison, the loss function calculation unit 503 calculates a plurality of types of losses, and then calculates an integrated loss by weighting and adding the calculated plurality of types of losses. Furthermore, the loss function calculation unit 503 updates model parameters of the encoder unit 501 and the decoder unit 502 based on the calculated integrated loss. In this way, a trained prediction model is generated. Note that the trained prediction model here includes a trained encoder unit and a trained decoder unit.

[0081] (3) Learning Unit Operation Example 1 Next, an operation example of the learning unit 112, specifically, an operation example of the prediction model among the prediction model and the loss function calculation unit, will be described. Fig. 6 is a first diagram showing an operation example of the learning unit.

[0082] FIG. 6(a) shows a state in which the "input data" of the learning data 300 is read out and input to the encoder unit 501, and the feature is output from the encoder unit 501. FIG. 6(a) shows a state in which the feature output from the encoder unit 501 is input to the decoder unit 502, and a start signal is input, and the decoder unit 502 outputs output data. In the example of FIG. 6(a), the output data output from the decoder unit 502 corresponds to the phase fraction at 800°C. The output data output from the decoder unit 502, specifically, the output data corresponding to the phase fraction at 800°C, is notified to the loss function calculation unit 503. The start signal is a signal for starting the operation of the decoder unit 502, and in this embodiment, an arbitrary value different from the output data output from the decoder unit 502 is input to the decoder unit 502 as the start signal. An arbitrary value different from the output data may be, for example, a value such as "000000".

[0083] 6(b) shows a state in which the feature quantity output from the encoder unit 501 is input to the decoder unit 502, and the correct data at 800° C. is input to the decoder unit 502, whereby the decoder unit 502 outputs output data. In the example of FIG. 6(b), the output data output by the decoder unit 502 corresponds to the phase fraction at 790° C. The output data output from the decoder unit 502, specifically, the output data corresponding to the phase fraction at 790° C., is notified to the loss function calculation unit 503.

[0084] Fig. 6(c) shows a state in which the feature quantity output from the encoder unit 501 is input to the decoder unit 502, and the correct data of the phase fraction up to 790°C is input to the decoder unit 502, whereby the decoder unit 502 outputs output data. However, due to space limitations, Fig. 6(c) shows only the correct data at 790°C. In reality, the correct data at 800°C and the correct data at 790°C are weighted and input to the decoder unit 502.

[0085] 6(c), the output data output by the decoder unit 502 corresponds to the phase fraction at 780° C. The output data output from the decoder unit 502, specifically, the output data corresponding to the phase fraction at 780° C., is notified to the loss function calculation unit 503.

[0086] Fig. 6(d) shows a state in which the features output from the encoder unit 501 are input to the decoder unit 502, and the correct data of the phase fraction up to 120°C is input to the decoder unit 502, whereby the decoder unit 502 outputs output data. However, due to space limitations, Fig. 6(d) shows only the correct data at 120°C. In reality, each correct data from 800°C to 120°C is weighted and input to the decoder unit 502.

[0087] 6(d), the output data output by the decoder unit 502 corresponds to the phase fraction at 110° C. The output data output by the decoder unit 502, specifically, the output data corresponding to the phase fraction at 110° C., is notified to the loss function calculation unit 503.

[0088] Fig. 6(e) shows a state in which the features output from the encoder unit 501 are input to the decoder unit 502, and the correct data of the phase fraction up to 110°C is input to the decoder unit 502, whereby the decoder unit 502 outputs output data. However, due to space limitations, Fig. 6(e) shows only the correct data at 110°C. In reality, each correct data from 800°C to 110°C is weighted and input to the decoder unit 502.

[0089] 6(e), the output data output by the decoder unit 502 corresponds to the phase fraction at 100° C. The output data output by the decoder unit 502, specifically, the output data corresponding to the phase fraction at 100° C., is notified to the loss function calculation unit 503.

[0090] (4) Learning unit operation example 2 Next, an operation example of the learning unit 112, specifically, an operation example of the loss function calculation unit 503 among the prediction model and loss function calculation units, will be described. FIG. 7 is a second diagram showing an operation example of the learning unit. As shown in FIG. 7, the loss function calculation unit 503 reads out "correct answer data" of the learning data 300. The example in FIG. 7 shows the phase fraction at each temperature in a specific temperature interval included in "phase fraction 1", The correct data is the phase fraction at 800℃. ·The correct data is the phase fraction at 790℃, The correct data is the phase fraction at 100°C. The figure shows the state in which the above information is read out.

[0091] In the example of FIG. 7, the decoder unit 502 sequentially outputs output data based on the feature amount output from the encoder unit 501 in response to the input of “material composition information 1.” As a result, the loss function calculation unit 503 receives the following: Output data: Phase fraction at 800°C Output data: phase fraction at 790°C, Output data: phase fraction at 100°C This shows how the

[0092] The example of Fig. 7 shows how the loss function calculation unit 503 calculates the integrated loss after comparing the phase fraction at each temperature from 100°C to 800°C, which is the correct answer data, with the phase fraction at each temperature from 100°C to 800°C, which is the output data. The example of Fig. 7 shows how the loss function calculation unit 503 updates the model parameters of the encoder unit 501 and the decoder unit 502 based on the calculated integrated loss. The functional configuration of the loss function calculation unit 503 for calculating the integrated loss will be described in detail below.

[0093] (5) Details of the functional configuration of the loss function calculation unit 8 is a diagram showing an example of a method for calculating an integrated loss by the loss function calculation unit. As shown in FIG. 8, the loss function calculation unit 503 has the following functions for calculating multiple types of losses: A phase fraction error calculation unit 801 at the formation temperature and disappearance temperature; ·Phase fraction error calculation unit 802, A phase fraction error (logarithmic value) calculation unit 803, A phase fraction error (difference value) calculation unit 804, Cross entropy error calculation unit 805, A weighting addition unit 806, has.

[0094] The phase fraction error calculation unit 801 at the generation temperature and disappearance temperature specifies the temperature at which the curve of each phase fraction becomes 0 among the correct answer data of the learning data. At the specified temperature, the phase fraction error calculation unit 801 at the generation temperature and disappearance temperature compares the phase fraction at each temperature from 100°C to 800°C, which is the correct answer data, with the phase fraction at each temperature from 100°C to 800°C, which is the output data. As a result, the phase fraction error calculation unit 801 at the generation temperature and disappearance temperature adds up the errors between the phase fraction at the temperature at which each phase specified based on the learning data is generated or disappeared and the phase fraction specified based on the learning data, for all phases. The phase fraction error calculation unit 801 at the generation temperature and disappearance temperature outputs a first addition result obtained by adding up all phases.

[0095] The phase fraction error calculation unit 802 compares the phase fraction at each temperature from 100° C. to 800° C., which is the correct data, with the phase fraction at each temperature from 100° C. to 800° C., which is the output data. As a result, the phase fraction error calculation unit 802 outputs a second addition result in which the phase fraction error at each temperature is added for a specific temperature interval.

[0096] The phase fraction error (logarithmic value) calculation unit 803 compares the phase fraction at each temperature from 100° C. to 800° C., which is the correct data, with the phase fraction at each temperature from 100° C. to 800° C., which is the output data. As a result, the phase fraction error (logarithmic value) calculation unit 803 outputs a third addition result in which the error in the logarithmic value of the phase fraction at each temperature within a specific temperature interval is added for the specific temperature interval.

[0097] The phase fraction error (difference value) calculation unit 804 compares the phase fraction at each temperature from 100° C. to 800° C., which is the correct data, with the phase fraction at each temperature from 100° C. to 800° C., which is the output data. As a result, the phase fraction error (difference value) calculation unit 804 outputs a fourth addition result in which the error in the phase fraction difference value between adjacent temperatures is added for a specific temperature interval.

[0098] The cross entropy error calculation unit 805 compares the phase fractions at each temperature from 100° C. to 800° C., which is the correct data, with the phase fractions at each temperature from 100° C. to 800° C., which is the output data. As a result, the cross entropy error calculation unit 805 outputs a fifth addition result in which the error in the phase fraction ratio at each temperature within a specific temperature interval is added for the specific temperature interval.

[0099] The weighted addition unit 806 calculates the integrated loss by weighting and adding the first to fifth addition results output from each unit from the phase fraction error calculation unit 801 to the cross entropy error calculation unit 805 at the formation temperature and disappearance temperature.

[0100] In this way, by calculating a plurality of types of losses and performing weighted addition, the loss function calculation unit 503 can process the loss between the output data and the correct answer data in a multifaceted manner. As a result, the learning unit 112 can appropriately update the model parameters when learning the prediction model.

[0101] <Learning process flow> Next, a description will be given of the flow of the learning process performed by the learning device 110. Fig. 9 is a flowchart showing the flow of the learning process.

[0102] In step S901, the learning data generating unit 111 receives, as an input of element information, the type of additive element to be added when generating a specific alloy, and the upper and lower limits of the content rate of each additive element.

[0103] In step S902, the learning data generating unit 111 randomly selects the content rate of each additional element within the constraints of the upper and lower limits, thereby determining a plurality of combinations of the content rates of each additional element.

[0104] In step S903, the learning data generating unit 111 operates a simulator for each of the material composition information indicated by the determined combinations, and calculates the phase fraction at each temperature within a specific temperature interval.

[0105] In step S904, the learning data generation unit 111 generates the learning data, and then stores the generated learning data in the learning data storage unit 113.

[0106] In step S 905 , the learning unit 112 inputs a start signal to the decoder unit 502 .

[0107] In step S906, the learning unit 112 reads out the "input data" of the learning data, and then inputs the read out "input data" to the prediction model.

[0108] In step S907, the learning unit 112 sequentially inputs the phase fractions at each temperature up to the i-th temperature of the “correct data” of the learning data, i.e., the correct data at each temperature up to the i-th temperature, and sequentially outputs the (i+1)th output data.

[0109] In step S908, the learning unit 112 determines whether or not all output data corresponding to the phase fraction at each temperature in the specific temperature range has been output. If it is determined in step S908 that there is a temperature for which output data corresponding to the phase fraction has not been output (NO in step S908), the process returns to step S907.

[0110] On the other hand, if it is determined in step S908 that all the output data corresponding to the phase fraction at each temperature within the specific temperature interval has been output (YES in step S908), the process proceeds to step S909.

[0111] In step S909, the learning unit 112 reads out the "correct data" of the learning data, and then compares the read out "correct data" with the output data.

[0112] In step S910, the learning unit 112 calculates a plurality of types of losses, and then calculates an integrated loss by weighting and adding the calculated plurality of types of losses. The learning unit 112 updates model parameters of the encoder unit 501 and the decoder unit 502 based on the calculated integrated loss.

[0113] In step S911, the learning unit 112 determines whether or not to continue the learning process. If it is determined in step S911 that the learning process is to be continued (YES in step S911), the process returns to step S906 and the same process is performed by reading the next "input data" of the learning data.

[0114] On the other hand, if it is determined in step S911 that the learning process is to be ended (NO in step S911), the process proceeds to step S912.

[0115] In step S912, the learning unit 112 stores the trained encoder unit and the trained decoder unit in the storage unit of the prediction device 120 and the storage unit of the prediction result analysis device 130 as trained prediction models.

[0116] <Details of each part of the prediction device> Next, the prediction section 122 of the prediction device 120 will be described in detail.

[0117] (1) Functional configuration of the prediction unit First, a description will be given of the functional configuration of the prediction unit 122. Fig. 10 is a first diagram showing an example of the functional configuration of the prediction unit.

[0118] 10, the prediction unit 122 includes a trained encoder unit 1001 and a trained decoder unit 1002 that have been trained by the training unit 112. The trained encoder unit 1001 and the trained decoder unit 1002 form a trained prediction model.

[0119] The trained encoder unit 1001 receives material composition information of the material to be predicted from the material composition input unit 121, calculates features, and then outputs the calculated features to the trained decoder unit 1002.

[0120] The trained decoder unit 1002 sequentially outputs prediction data by inputting the feature values ​​output from the trained encoder unit 1001. Specifically, the trained decoder unit 1002 outputs prediction data at the (i+1)th temperature using prediction data at each temperature up to the i-th temperature. This enables the trained decoder unit 1002 to predict the phase fraction at each temperature within a specific temperature interval based on the feature values ​​output from the trained encoder unit 1001 by inputting material composition information of the material to be predicted.

[0121] (2) Example of operation of the prediction unit Next, an example of the operation of the prediction unit 122 will be described. Fig. 11 is a first diagram showing an example of the operation of the prediction unit. Of these, Fig. 11(a) shows a state in which material composition information of a material to be predicted is input to the trained encoder unit 1001, and a feature is output from the trained encoder unit 1001. Fig. 11(a) also shows a state in which the feature output from the trained encoder unit 1001 is input to the trained decoder unit 1002, and a start signal is input to the trained decoder unit 1002, and a phase fraction at 800°C is output.

[0122] 11(b) shows how the feature quantity output from the trained encoder unit 1001 is input to the trained decoder unit 1002, and how the phase fraction at 800° C., which is the predicted data, is input to the trained decoder unit 1002. As a result, the trained decoder unit 1002 outputs the phase fraction at 790° C. as the predicted data.

[0123] FIG. 11(c) shows a state in which the feature quantity output from the trained encoder unit 1001 is input to the trained decoder unit 1002, and the phase fraction up to 790°C, which is the predicted data, is input to the trained decoder unit 1002. Note that, due to space limitations, FIG. 11(c) shows only the phase fraction at 790°C. In reality, the phase fraction at 800°C and the phase fraction at 790°C are weighted and input to the trained decoder unit 1002. As a result, the trained decoder unit 1002 outputs the phase fraction at 780°C as the predicted data.

[0124] FIG. 11(d) shows how the feature values ​​output from the trained encoder unit 1001 are input to the trained decoder unit 1002, and how the phase fraction up to 120° C., which is the predicted data, is input to the trained decoder unit 1002. Note that, due to space limitations, FIG. 11(d) only shows the phase fraction at 120° C. In reality, the phase fractions from 800° C. to 120° C. are weighted and input to the trained decoder unit 1002. As a result, the trained decoder unit 1002 outputs the phase fraction at 110° C. as the predicted data.

[0125] FIG. 11(e) shows a state in which the feature values ​​output from the trained encoder unit 1001 are input to the trained decoder unit 1002, and the phase fraction up to 110° C., which is the predicted data, is input to the trained decoder unit 1002. Note that, due to space limitations, only the phase fraction at 110° C. is shown in FIG. 11(e). In reality, the phase fractions from 800° C. to 110° C. are weighted and input to the trained decoder unit 1002. As a result, the trained decoder unit 1002 outputs the phase fraction at 100° C. as the predicted data. Note that the example of FIG. 11(e) also shows the phase fraction at 100° C., which is the predicted data output from the trained decoder unit 1002, as a result of inputting the feature values ​​and the phase fraction up to 110° C., which is the predicted data.

[0126] <Prediction process flow> Next, a description will be given of the flow of prediction processing by the prediction device 120. Fig. 12 is a flowchart showing the flow of the prediction processing.

[0127] In step S1201, the material composition input unit 121 receives input of material composition information of the material to be predicted.

[0128] In step S1202, 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 specific temperature interval.

[0129] In step S1203, the display unit 123 displays the predicted phase fraction at each temperature.

[0130] <Prediction accuracy of phase fraction> Next, the prediction accuracy of the phase fraction at each temperature in a specific temperature interval predicted by the prediction unit 122 will be described. FIG. 13 is a diagram for explaining the prediction accuracy of the phase fraction. Each of FIGS. 13(a) to 13(c) shows prediction data when the content rate of each additive element constituting "6013", which is the name of a known aluminum alloy standard, is input as material composition information of the material to be predicted. In this embodiment, the correct answer data shown by reference numeral 311 in FIG. 3 is used to evaluate the prediction accuracy. To calculate the loss from the correct answer data, the mean squared logarithmic error (MSLE) loss was used.

[0131] FIG. 13(a) shows predicted data when a fully connected multilayer neural network is applied to the prediction model as a comparative example. In the case of FIG. 13(a), the MSLE loss with respect to the correct data is 3.79×10 -4 It becomes.

[0132] Figure 13(b) shows the predicted data when applying the attention mechanism to the prediction model, Seq2Seq. In the case of Figure 13(b), the MSLE loss with the ground truth data is 5.10 × 10 -5 It becomes.

[0133] Figure 13(c) shows the predicted data when Transformer is applied to the prediction model. In the case of Figure 13(c), the MSLE loss with the ground truth data is 2.45 × 10 -5 It becomes.

[0134] In this way, when a fully connected multilayer neural network was applied as a prediction model to calculate the phase fraction at low cost, the estimation accuracy was low in some temperature ranges, for example, from 400°C to 600°C.

[0135] On the other hand, when we applied Seq2Seq with an Attention mechanism as a prediction model that realizes phase fraction calculation at low cost, we were able to significantly improve the loss.Furthermore, when we applied Transformer, we were able to reproduce phase fractions at roughly the same level as the correct data.

[0136] That is, when predicting the phase fraction over a specific temperature range based on material composition information, - Processing the phase fraction at each temperature as time series data using a recurrent network, and - Using multiple types of loss functions to handle losses in a multifaceted way; This can improve the prediction accuracy.

[0137] <Details of the prediction result analysis device> Next, as details of the prediction result analysis device 130, specific examples of the processing of the material composition input unit 131, the baseline composition input unit 132, the intermediate composition generation unit 133, and the integral gradient calculation unit 135, and an example of the functional configuration and operation of the prediction unit 134 will be described.

[0138] (1) Specific examples of processing in the material composition input section, baseline composition input section, and intermediate composition generation section First, specific examples of the processing of the material composition input unit 131, the baseline composition input unit 132, and the intermediate composition generating unit 133 among the units of the prediction result analysis device 130 will be described. Fig. 14 is a first diagram showing a specific example of the processing of the prediction result analysis device.

[0139] 14, the material composition input unit 131 accepts input of "material composition information X" as material composition information of the material to be predicted. The baseline composition input unit 132 accepts input of the number "N" of baseline composition information, and also accepts input of multiple baseline composition information, specifically, N pieces of baseline composition information, "baseline composition information 1" to "baseline composition information N".

[0140] The intermediate composition generating unit 133 generates M pieces of intermediate composition information by dividing the material composition information of the material to be predicted and each of the N pieces of baseline composition information by a predetermined number of steps=M. In the example of FIG. "Intermediate composition information Xto1_1" to "Intermediate composition information Xto1_M" are generated by dividing "Material composition information X" and "Baseline composition information 1" by the number of steps = M. "Intermediate composition information Xto2_1" to "Intermediate composition information Xto2_M" are generated by dividing "Material composition information X" and "Baseline composition information 2" by the number of steps = M. ... "Intermediate composition information XtoN_1" to "Intermediate composition information XtoN_M" are generated by dividing "material composition information X" and "baseline composition information N" by the number of steps = M. This shows that.

[0141] (2) Functional configuration of the prediction unit Next, a detailed description will be given of the functional configuration of the prediction unit 134. Fig. 15 is a second diagram showing an example of the functional configuration of the prediction unit.

[0142] 15, the prediction unit 134 includes a trained encoder unit 1001 and a trained decoder unit 1002 that have been trained by the training unit 112. The trained encoder unit 1001 and the trained decoder unit 1002 form a trained prediction model.

[0143] The functions of the trained encoder unit 1001 and the trained decoder unit 1002 are the same as those of the trained encoder unit 1001 and the trained decoder unit 1002 in Fig. 10, except that intermediate composition information is input instead of material composition information of the material to be predicted. Therefore, a description thereof will be omitted here.

[0144] 15, the prediction unit 134 further includes a phase designation unit 1501 and a sensitivity characteristic calculation unit 1502. The phase designation unit 1501 and the sensitivity characteristic calculation unit 1502 form an input sensitivity analysis unit.

[0145] The phase designation unit 1501 accepts input of the phase of the analysis target designated by the material designer. The phase of the analysis target designated by the material designer refers to a phase that the material designer wants to know in advance how much it will change at each temperature when each component included in the material composition is changed. The phase designation unit 1501 notifies the sensitivity characteristic calculation unit 1502 of the phase of the analysis target that has been input.

[0146] When the predicted data at the (i+1)th temperature is output, the sensitivity characteristic calculation unit 1502 derives a partial differential value by partially differentiating the predicted value for the phase notified by the phase designation unit 1501 using the intermediate composition information notified by the intermediate composition generation unit 133.

[0147] Specifically, the sensitivity characteristic calculation unit 1502 Partial differential values ​​in each layer calculated by processing in each layer in the trained encoder unit 1001 during the period from when intermediate composition information is input until features are output, and Partial differential values ​​in each layer calculated by processing in each layer in the trained decoder unit 1002 during the time from when a feature is input until a predicted value for a specified phase at the (i+1)th temperature is output; and By using this, "a partial differential value obtained by partially differentiating the value at the (i+1)th temperature predicted for the specified analysis phase using intermediate composition information" is derived.

[0148] In addition, the sensitivity characteristic calculation unit 1502 notifies the integral gradient calculation unit 135 of “the partial differential value obtained by partially differentiating the value at the (i+1)th temperature predicted for the specified analysis target phase using the intermediate composition information.”

[0149] (3) Example of operation of the prediction unit Next, a specific example of the process of the prediction unit 134 will be described. Fig. 16A to Fig. 16C are second to fourth diagrams showing an operation example of the prediction unit. Among these, Fig. 16A (a) shows a state in which intermediate composition information generated based on material composition information of a material to be predicted and baseline composition information is input to the trained encoder unit 1001, and a feature amount is output from the trained encoder unit 1001. Fig. 16A (a) also shows a state in which the feature amount output from the trained encoder unit 1001 is input to the trained decoder unit 1002, and a start signal is input to the trained decoder unit 1002, and a phase fraction at 800°C is output.

[0150] Moreover, (a) of FIG. 16A shows how, when the phase fraction at 800°C is output as prediction data, the "partial differential value obtained by partially differentiating the predicted value for the specified phase using the intermediate composition information" is derived for each component at 800°C.

[0151] 16A(b) shows a state in which the feature quantity output from the trained encoder unit 1001 is input to the trained decoder unit 1002, and the phase fraction at 800° C., which is the predicted data, is input to the trained decoder unit 1002. As a result, the trained decoder unit 1002 outputs the phase fraction at 790° C. as the predicted data.

[0152] Moreover, (b) of FIG. 16A shows how, when the phase fraction at 790°C is output as prediction data, the "partial differential value obtained by partially differentiating the predicted value for the specified phase using the intermediate composition information" is derived for each component at 790°C.

[0153] FIG. 16B (a) shows a state in which the feature quantity output from the trained encoder unit 1001 is input to the trained decoder unit 1002, and the phase fraction up to 790°C, which is the predicted data, is input to the trained decoder unit 1002. Note that, due to space limitations, FIG. 16B (a) shows only the phase fraction at 790°C. In reality, the phase fraction at 800°C and the phase fraction at 790°C are weighted and input to the trained decoder unit 1002. As a result, the trained decoder unit 1002 outputs the phase fraction at 780°C as the predicted data.

[0154] Moreover, (a) of FIG. 16B shows how, when the phase fraction at 780°C is output as prediction data, the "partial differential value obtained by partially differentiating the predicted value for the specified phase using the intermediate composition information" is derived for each component at 780°C.

[0155] FIG. 16B(b) shows a state in which the feature quantity output from the trained encoder unit 1001 is input to the trained decoder unit 1002, and the phase fraction up to 120° C., which is the predicted data, is input to the trained decoder unit 1002. Note that, due to space limitations, FIG. 16B(b) shows only the phase fraction at 120° C. In reality, the phase fractions from 800° C. to 120° C. are weighted and input to the trained decoder unit 1002. As a result, the trained decoder unit 1002 outputs the phase fraction at 110° C. as the predicted data.

[0156] Moreover, (b) of FIG. 16B shows how, when the phase fraction at 110°C is output as prediction data, the "partial differential value obtained by partially differentiating the predicted value for the specified phase using the intermediate composition information" is derived for each component at 110°C.

[0157] FIG. 16C shows a state in which the feature values ​​output from the trained encoder unit 1001 are input to the trained decoder unit 1002, and the phase fractions up to 110° C., which are the predicted data, are input to the trained decoder unit 1002. Note that, due to space limitations, only the phase fraction at 110° C. is shown in FIG. 16C. In reality, the phase fractions from 800° C. to 110° C. are weighted and input to the trained decoder unit 1002. As a result, the trained decoder unit 1002 outputs the phase fraction at 100° C. as the predicted data. Note that the example of FIG. 16C also shows the phase fraction at 100° C., which is the predicted data output from the trained decoder unit 1002, as a result of inputting the feature values ​​and the phase fractions up to 110° C., which are the predicted data.

[0158] Moreover, FIG. 16C shows how, when the phase fraction at 100°C is output as prediction data, the "partial differential value obtained by partially differentiating the predicted value for the specified phase using the intermediate composition information" is derived for each component at 100°C.

[0159] (4) Specific examples of the processing of the prediction unit and integral gradient calculation unit Next, a specific example of the processing by the prediction unit 134 and the integral gradient calculation unit 135 among the units included in the prediction result analysis device 130 will be described. Fig. 17 is a second diagram showing a specific example of the processing by the prediction result analysis device.

[0160] 17, the prediction unit 134 inputs the M pieces of intermediate composition information notified from the intermediate composition generation unit 133 into the trained prediction model, thereby sequentially predicting the M phase fractions at each temperature within a specific temperature interval. The prediction unit 134 also derives M partial differential values ​​for each temperature and each component for the phase to be analyzed.

[0161] According to the example of FIG. 17, the prediction unit 134 By inputting M pieces of intermediate composition information, "Intermediate composition information Xto1_1" to "Intermediate composition information Xto1_M", M phase fractions in the case of "Baseline composition information 1" are predicted, and M partial differential values ​​are derived for each temperature and each component for the phase to be analyzed. By inputting M pieces of intermediate composition information, "Intermediate composition information Xto2_1" to "Intermediate composition information Xto2_M", M phase fractions in the case of "Baseline composition information 2" are predicted, and M partial differential values ​​are derived for each temperature and each component for the phase to be analyzed. ... By inputting M pieces of intermediate composition information, "Intermediate composition information XtoN_1" to "Intermediate composition information XtoN_M", M phase fractions for the "Baseline composition information N" are predicted, and M partial differential values ​​are derived for each temperature and component for the phase to be analyzed.

[0162] As shown in FIG. 17, the integral gradient calculation unit 135 calculates the integral gradient of each temperature and each component by using M partial differential values ​​derived by the prediction unit 134 for each temperature and each component.

[0163] In the example of FIG. 17, the integral gradient calculation unit 135 The partial differential values ​​of each temperature and each component derived from "intermediate composition information Xto1_M" were obtained. The partial differential values ​​of each temperature and each component derived for "intermediate composition information Xto1_M-1" were obtained. ... The partial differential values ​​of each temperature and each component derived for "intermediate composition information Xto1_1" were obtained. The integral gradient 1710_1 is calculated by integrating M partial differential values ​​for each temperature and each component using the composition weights calculated in advance and then averaging them. This shows that.

[0164] Similarly, in the example of FIG. 17, the integral gradient calculation unit 135 The partial differential values ​​of each temperature and each component derived from "intermediate composition information Xto2_M" were obtained. The partial differential values ​​of each temperature and each component derived for "intermediate composition information Xto2_M-1" were obtained. ... The partial differential values ​​of each temperature and each component derived for "intermediate composition information Xto2_1" were obtained. The integral gradient 1710_2 is calculated by integrating M partial differential values ​​for each temperature and each component using the composition weights calculated in advance and then averaging them. This shows that.

[0165] Similarly, in the example of FIG. 17, the integral gradient calculation unit 135 The partial differential values ​​of each temperature and each component derived from "intermediate composition information XtoN_M" were obtained. The partial differential values ​​of each temperature and each component derived for "intermediate composition information XtoN_M-1" were obtained. ... The partial differential values ​​of each temperature and each component derived for "intermediate composition information XtoN_1" were obtained. The integral gradient 1710_N was calculated by integrating M partial differential values ​​for each temperature and each component using the composition weights calculated in advance and then averaging them. This shows that.

[0166] Furthermore, the example of Figure 17 shows how the integral gradient calculation unit 135 calculates the final integral gradient for the phase being analyzed, i.e., the averaged integral gradient, by calculating the average value of multiple integral gradients 1710_1 to 1710_N for each temperature and component.

[0167] <Flow of prediction result analysis process> Next, a description will be given of the flow of the prediction result analysis process performed by the prediction result analysis device 130. Fig. 18 is a flowchart showing the flow of the prediction result analysis process.

[0168] In step S1801, the material composition input unit 131 receives input of material composition information of the material to be predicted.

[0169] In step S1802, the baseline composition input unit 132 accepts input of baseline composition information. A detailed flowchart of the baseline composition information input process will be described later.

[0170] In step S1803, the prediction unit 134 receives an input of the phase to be analyzed.

[0171] In step S1804, the intermediate composition generating unit 133 generates intermediate composition information based on the material composition information of the material to be predicted and the baseline composition information. The predicting unit 134 predicts the phase fraction based on the generated intermediate composition information, and derives partial differential values ​​for each temperature and each component for the phase to be analyzed. The integral gradient calculating unit 135 calculates the integral gradient for each temperature and each component by integrating the derived partial differential values ​​for each temperature and each component. A detailed flowchart of the integral gradient calculation process by the intermediate composition generating unit 133, the predicting unit 134, and the integral gradient calculating unit 135 will be described later.

[0172] In step S1805, the display unit 136 displays a heat map of the integral gradient for the phase to be analyzed.

[0173] In this way, the prediction result analysis device 130 makes it possible to grasp the degree of influence of the content of each additive element on the phase fraction simply by inputting multiple baseline composition information and the phase to be analyzed.

[0174] In other words, it is no longer necessary to predict the phase fraction within a specific temperature range while sequentially increasing or decreasing the content of all additive elements, making it easier to grasp the degree of effect of the content of each additive element on the phase fraction.

[0175] <Detailed flow of baseline composition information input process> Next, a detailed flow of the baseline composition information input process (step S1802) will be described below. Fig. 19 is a flowchart showing the flow of the baseline composition information input process.

[0176] In step S1901, the baseline composition input unit 132 receives a setting for the number (=N) of baseline composition information from a material designer.

[0177] In step S1902, the baseline composition input unit 132 sets a counter i, which counts the number of pieces of baseline composition information, to "1."

[0178] In step S1903, the baseline composition input unit 132 receives input of the i-th baseline composition information from the material designer.

[0179] In step S1904, the baseline composition input unit 132 determines whether or not the counter i has reached N. If it is determined in step S1904 that the counter i has not reached N (NO in step S1904), the process proceeds to step S1905.

[0180] In step S1905, the baseline composition input unit 132 increments the counter i, and then returns to step S1903.

[0181] On the other hand, if it is determined in step S1905 that the counter i has reached N (YES in step S1905), the baseline composition information input process ends.

[0182] <Detailed flow of integral gradient calculation process> Next, a detailed flow of the integral gradient calculation process (step S1804) will be described below. Fig. 20 is a flowchart showing the flow of the integral gradient calculation process.

[0183] In step S2001, the intermediate composition generating unit 133 sets "1" to a counter i that counts the number of pieces of baseline composition information.

[0184] In step S2002, the intermediate composition generating unit 133 generates M pieces of intermediate composition information by dividing the interval between the material composition information of the material to be predicted and the i-th baseline composition information by the number of steps=M.

[0185] In step S2003, the intermediate composition generating unit 133 sets "1" to a counter j that counts the number of pieces of intermediate composition information generated.

[0186] In step S2004, the intermediate composition generating unit 133 calculates the composition weight of the j-th intermediate composition information. The intermediate composition generating unit 133 calculates the composition weight based on the difference between the material composition information of the material to be predicted and the j-th intermediate composition information.

[0187] In step S2005, the prediction unit 134 predicts the phase fraction at each temperature within a specific temperature interval by inputting the j-th intermediate composition information into the trained prediction model.

[0188] In step S2006, the prediction unit 134 performs partial differentiation on the phase to be analyzed using the j-th intermediate composition information, thereby deriving partial differential values ​​for each temperature and each component.

[0189] In step S2007, the integral gradient calculation unit 135 corrects the partial differential values ​​derived for each temperature and each component by using the composition weights.

[0190] In step S2008, the intermediate composition generating unit 133 judges whether or not the counter j that counts the number of intermediate composition information has reached M. In step S2008, if it is judged that the counter j has not reached M (NO in step S2008), the process proceeds to step S2009.

[0191] In step S2009, the intermediate composition generating unit 133 increments the counter j, and then returns to step S2004.

[0192] On the other hand, if it is determined in step S2008 that the counter j has reached M (YES in step S2008), the process proceeds to step S2010.

[0193] In step S2010, the integral gradient calculation unit 135 integrates the M corrected partial differential values ​​derived for each temperature and each component, and then averages them by dividing them by M. In this way, the integral gradient calculation unit 135 calculates the integral gradient of each temperature and each component corresponding to the i-th baseline information.

[0194] In step S2011, the intermediate composition generating unit 133 judges whether or not the counter i for counting the number of baseline composition information has reached N. In step S2011, if it is judged that the counter i has not reached N (NO in step S2011), the process proceeds to step S2012.

[0195] In step S2012, the intermediate composition generating unit 133 increments the counter i, and then returns to step S2002.

[0196] On the other hand, if it is determined in step S2011 that the counter i has reached N (YES in step S2011), the process proceeds to step S2013.

[0197] In step S2013, the N integral gradients for each temperature and each component are integrated for each temperature and each component, and then averaged by dividing by N. This allows the integral gradient calculation unit 135 to calculate the final integral gradient, that is, the averaged integral gradient, for the material composition information of the material to be predicted.

[0198] <Example of heat map> Next, a specific example of a heat map of integral gradient calculated by the prediction result analysis process (FIG. 18) by the prediction result analysis device 130 will be described with reference to FIGS. 21 and 22. FIG. 21 is a diagram showing an example of the material composition of the material to be predicted used in the prediction result analysis process and the predicted phase fraction. FIG. 22 is a diagram showing an example of a heat map displayed in a display mode according to the integral gradient.

[0199] As shown in FIG. 21, when the material composition information of the prediction target material is the material composition information shown by the reference numeral 2110, the prediction unit 122 predicts the reference numeral 2120 as the phase fraction at each temperature from 100° C. to 800° C. In contrast, when the material design designer predicts Al as the phase to be analyzed, 15 When the prediction result analysis device 130 calculates a heat map of the integral gradient by setting (Mn,Fe)3Si2, for example, the heat maps shown in FIGS. 22(a) and (b) are obtained.

[0200] 22(a) and 22(b) show heat maps in which the number of pieces of baseline composition information (=N) is set to "1." However, it is assumed that different pieces of baseline composition information are input in FIG. 22(a) and FIG. 22(b).

[0201] By referring to Figure 22, material designers can predict the effect of each additive element (Si, Fe, Cu, Mn, Mg, Cr, Zn, Ti) in the material to be predicted. The phase to be analyzed is Al 15 To what extent does each added element affect (Mn,Fe)3Si2? At what temperature does it have an effect? can be recognized.

[0202] In this way, according to the prediction result analysis device 130, a heat map is displayed for the analyzed phase in a display mode according to the integral gradient calculated for each temperature and each component, so that the designer can easily grasp the degree of influence of the content of each added element on the phase fraction.

[0203] In other words, compared to comparing the phase fractions before and after changing the content, the designer can more easily grasp the degree of effect of the content of each added element on the phase fraction.

[0204] In the examples of Figures 22(a) and (b), the number of baseline composition information pieces is set to "1". However, by further increasing the number of baseline composition information pieces, the degree of influence of each added element on the phase being analyzed can be calculated with greater accuracy.

[0205] In addition, in the examples of Figures 22(a) and (b), no mention is made of the number of steps = M. However, by further increasing the number of steps = M, the degree of influence of each added element on the phase to be analyzed can be calculated with even greater accuracy.

[0206] <Summary> As is clear from the above description, the prediction result analysis device 130 according to the first embodiment: The device has a memory unit that stores a trained prediction model trained using training data in which the material composition of the training target material is associated with the phase fraction of the training target material at each temperature within a specific temperature range. - Split the material composition of the material to be forecasted and the material composition of the baseline material to generate multiple intermediate material compositions. By inputting the material compositions of multiple intermediate materials into a trained prediction model, the phase fractions of multiple intermediate materials at each temperature within a specific temperature range are predicted. When predicting the phase fraction, partial differential values ​​are calculated by partially differentiating the predicted value for the phase to be analyzed at each temperature using each of the material compositions of multiple intermediate materials. The calculated partial differential value is integrated for each component contained in the material composition of the plurality of intermediate materials at each temperature to calculate the integral gradient for each temperature and each component. In a heat map in which each component is arranged on the first axis and each temperature is arranged on the second axis, each region identified based on each component and each temperature is displayed in a display mode according to the integral gradient.

[0207] As a result, according to the first embodiment, the designer can easily grasp the degree of effect on the phase fraction due to a change in the material composition.

[0208] [Second embodiment] In the above first embodiment, the color scheme of the heat map is the color scheme shown in (a) and (b) of Fig. 22, but the heat map may have a color scheme other than the color scheme shown in (a) and (b) of Fig. 22. Alternatively, the heat map may be expressed in a display mode other than the color scheme according to the integral gradient.

[0209] In the above first embodiment, a case has been described in which the calculated integral gradient is expressed using a heat map so that the designer can easily understand which temperature interval phase fraction can be affected by increasing or decreasing the content of which additive element. However, the method of expressing the calculated integral gradient is not limited to the heat map. For example, in a case where it is desired to understand which temperature in a specific temperature interval has a larger degree of influence, a graph showing the transition of the degree of influence due to the difference in temperature may be displayed instead of the heat map.

[0210] In the above first embodiment, the alloy composition showing the content of additional elements, specifically other metal elements or nonmetal elements, to the metal elements contained in the alloy is described as a specific example of the material composition, but the material composition is not limited to the alloy composition. For example, it may be a chemical composition showing the content of each chemical component contained in a substance other than the alloy.

[0211] In the above first embodiment, the trained prediction model is described as applying either the architecture of Seq2Seq with an Attention mechanism or Transformer. However, the trained prediction model is not limited to a model applying these architectures, and may be a model applying other architectures as long as the architecture is capable of processing time series data, which is data for each specific time interval. Specifically, the trained prediction model may be a model applying an architecture such as RNN (Recurrent Neural Network), Bidirectional RNN, Seq2Seq, etc. Alternatively, the trained prediction model may be a model applying an architecture such as GRU (Gated Recurrent Unit), LSTM (Long Short Term Memory), etc.

[0212] In the above first embodiment, the learning device 110, the prediction device 120, and the prediction result analysis device 130 are configured as separate entities, but the learning device 110, the prediction device 120, and the prediction result analysis device 130 may be configured as a single device. Alternatively, any two of the learning device 110, the prediction device 120, and the prediction result analysis device 130 may be configured as a single device.

[0213] The present disclosure is not limited to the configurations described in the above embodiments, combinations with other elements, etc. These points can be changed without departing from the spirit of the present disclosure, and can be appropriately determined according to the application form.

[0214] This application claims priority based on Japanese Patent Application No. 2023-038608, filed on March 13, 2023, the entire contents of which are incorporated herein by reference. [Explanation of symbols]

[0215] 100: Prediction system 110: Learning device 111: Learning data generation unit 112: Learning Department 120: Prediction device 121: Material composition input section 122: Prediction section 123: Display section 130: Prediction result analysis device 131: Material composition input section 132: Baseline composition input section 133: Intermediate composition generation section 134: Prediction Department 135: Integral gradient calculation unit 136:Display section 300: Training data 401: Element information input section 402: Combination determination unit 403: Simulation Department 404: Storage control unit 501: Encoder section 502: Decoder section 503: Loss function calculation unit 801: Phase fraction error calculation section at formation temperature and disappearance temperature 802: Phase fraction error calculation section 803: Phase fraction error (logarithmic value) calculation section 804: Phase fraction error (difference value) calculation unit 805: Cross entropy error calculation unit 806: Weighting addition unit 1001: Trained encoder 1002: Trained decoder part 1501: Phase specification section 1502: Sensitivity characteristic calculation unit

Claims

1. a memory unit that stores a trained model that is trained using training data in which a material composition of a training target material is associated with a phase fraction of the training target material at each temperature within a specific temperature interval, the trained model predicting a phase fraction at an (i+1)th temperature using phase fractions predicted by the trained model for each temperature up to an i-th temperature (i is an integer equal to or greater than 1) within the specific temperature interval; an intermediate composition generating unit that divides a material composition of a material to be predicted and a material composition of a baseline material to generate material compositions of a plurality of intermediate materials; a prediction unit that, when predicting a phase fraction at each temperature of each of the plurality of intermediate materials in the specific temperature range by inputting material compositions of the plurality of intermediate materials into the trained model, calculates a partial differential value by partially differentiating the predicted value for a pre-specified phase at each temperature using each of the material compositions of the plurality of intermediate materials; an integral gradient calculation unit that calculates an integral gradient for each temperature and each component by integrating the calculated partial differential value for each component included in the material composition of the intermediate materials at each temperature; a display unit that displays, in a heat map in which each component is arranged on a first axis and each temperature is arranged on a second axis, each region specified based on each component and each temperature in a display mode corresponding to the integral gradient; A prediction result analysis device having the above configuration.

2. 2. The prediction result analysis device according to claim 1, wherein the integral gradient calculation unit calculates the integral gradient for each temperature and each component by integrating the calculated partial differential values ​​for each component contained in the material compositions of the intermediate materials at each temperature and then averaging the integral gradients.

3. 3. The prediction result analysis device according to claim 2, wherein when integrating the calculated partial differential value for each component contained in a material composition of the plurality of intermediate materials at each temperature, the integral gradient calculation unit integrates the partial differential value using a composition weight calculated in advance for each of the plurality of intermediate materials.

4. The prediction result analysis device according to claim 3 , wherein the integral gradient calculation unit calculates the composition weight based on a difference between each of the material compositions of the plurality of intermediate materials and a material composition of the prediction target material.

5. The intermediate composition generating unit generates material compositions of the intermediate materials for each of a plurality of baseline material compositions, The integral gradient calculation unit calculates the integral gradient of each temperature and each component for each material composition of the plurality of baseline materials; 5. The prediction result analysis device according to claim 1, wherein the display unit displays the heat map in a display mode according to an averaged integral gradient obtained by averaging a plurality of the integral gradients for each temperature and each component.

6. The prediction result analysis device according to any one of claims 1 to 5, wherein an architecture capable of processing time series data, which is data for each specific time interval, is applied to the trained model, and the trained model predicts phase fraction for each specific temperature interval based on the material composition of the material to be predicted.

7. The prediction result analysis device according to claim 6 , wherein the trained model is any one of RNN, Bidirectional RNN, Seq2Seq, Seq2Seq with an Attention mechanism, GRU, LSTM, and Transformer.

8. The trained model is an encoder unit that outputs a feature quantity in response to an input of a material composition of a material to be predicted; a decoder unit which predicts the phase fraction at the (i+1)th temperature by inputting the output feature amount and the predicted phase fractions at the temperatures up to the i-th temperature; The prediction result analysis device according to claim 7 , further comprising:

9. The prediction result analysis device according to claim 1 , wherein the phase fraction is a phase fraction in a thermodynamic equilibrium state.

10. An analysis method, comprising: a trained model that is trained using training data in which a material composition of a training target material is associated with a phase fraction at each temperature of the training target material within a specific temperature interval; the trained model predicts a phase fraction at an (i+1)th temperature using phase fractions predicted by the trained model for temperatures up to an i-th temperature (i is an integer of 1 or more) within the specific temperature interval, the method comprising: The computer, Dividing the material composition of the material to be predicted and the material composition of the baseline material to generate a plurality of intermediate material compositions; a step of inputting material compositions of the intermediate materials into the trained model to predict phase fractions at each temperature of the intermediate materials in the specific temperature range, partially differentiating the predicted values ​​for a pre-specified phase at each temperature using the material compositions of the intermediate materials to calculate partial differential values; calculating an integral gradient for each temperature and each component by integrating the calculated partial differential value for each component contained in the material composition of the intermediate materials at each temperature; a step of displaying, in a heat map in which each component is arranged on a first axis and each temperature is arranged on a second axis, each region specified based on each component and each temperature in a display mode according to the integral gradient; The analysis method to be performed.

11. A prediction program for causing a computer to execute a trained model that is trained using training data in which a material composition of a training target material is associated with a phase fraction at each temperature of the training target material within a specific temperature interval, the trained model predicting a phase fraction at an (i+1)th temperature using phase fractions predicted by the trained model for each temperature up to an i-th temperature (i is an integer of 1 or more) within the specific temperature interval, The computer includes: Dividing the material composition of the material to be predicted and the material composition of the baseline material to generate a plurality of intermediate material compositions; a step of inputting material compositions of the intermediate materials into the trained model to predict phase fractions at each temperature of the intermediate materials in the specific temperature range, partially differentiating the predicted values ​​for a pre-specified phase at each temperature using the material compositions of the intermediate materials to calculate partial differential values; calculating an integral gradient for each temperature and each component by integrating the calculated partial differential value for each component contained in the material composition of the intermediate materials at each temperature; a step of displaying, in a heat map in which each component is arranged on a first axis and each temperature is arranged on a second axis, each region specified based on each component and each temperature in a display mode according to the integral gradient; An analysis program for executing the above.

Images