Visualization method for training data, learning data visualization device, and program

Abstract

To provide a visualization method for training data capable of intuitive comprehension of data used for learning an AI surrogate model for predicting the temperature distribution of an object, such as a substrate.SOLUTION: A visualization method for training data includes the following steps of: storing multiple pieces of training data used to train an AI model that predicts the temperature distribution on an object with components placed thereon in a memory; reading multiple training data corresponding to the specified AI model from the memory based on the specified the AI model; and outputting the multiple pieces of training data read out from the memory on a display device in a predetermined graph form.SELECTED DRAWING: Figure 7

Description

[Technical Field]

[0001] The present disclosure relates to a training data visualization method, a training data visualization device, and a program. [Background technology]

[0002] Patent Document 1 discloses a system that supports the decision on whether to accept or reject a model. In this system, in the information registered by the system, a learning model is associated with a dataset, which is one or more dataset elements that serve as input to the learning model, and a filter for the dataset is associated with the dataset. The system displays at least a portion of the information associated with the learning model to be viewed and information indicating the evaluation result of the learning model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2020-166681 Summary of the Invention [Problem to be solved by the invention]

[0004] In recent years, there has been a demand for faster processing speeds and more sophisticated functionality in electronic devices. However, control boards and other components built into the housings of electronic devices are often equipped with numerous heat-generating elements, such as electronic components that generate heat when power is supplied. This heat can potentially slow down processing speeds. Therefore, during the design phase of electronic device development, there is a need to accurately evaluate the heat generated on a board based on the placement of electronic components to be mounted on the board. The temperature distribution on a board with heat-generating elements can be predicted using simulation techniques such as Computational Fluid Dynamics (CFD), but when a board contains multiple heat-generating elements, the analysis can take an extremely long time, from tens of minutes to tens of hours.

[0005] As a replacement for CFD, AI surrogate models have emerged that use artificial intelligence (AI) to learn from the results of simulations such as CFD and make predictions. However, there is currently no way for anyone other than the creator of the AI ​​surrogate model to know the specific characteristics of the AI ​​surrogate model, such as the data it was trained on. Because the processing of AI surrogate models differs from CFD calculations, which solve predetermined physical equations for the target object to be predicted, users of AI surrogate models often find it difficult to objectively state that their output is highly reliable.

[0006] The present disclosure was devised in consideration of the above-described conventional situation, and aims to present data used in training an AI surrogate model that predicts the temperature distribution of an object such as a substrate in a manner that allows for intuitive understanding. [Means for solving the problem]

[0007] The present disclosure provides a learning data visualization method that stores in a memory a plurality of learning data used during learning of an AI model that predicts the temperature distribution of an object on which components are placed, reads from the memory the plurality of learning data corresponding to the specified AI model based on the specification of the AI ​​model, and outputs the plurality of learning data read from the memory to a display device in a predetermined graph format.

[0008] The present disclosure also provides a learning data visualization device comprising a processor and a memory that stores a plurality of learning data used during learning of an AI model that predicts the temperature distribution of an object on which components are placed, wherein the processor cooperates with the memory to read from the memory the plurality of learning data corresponding to the specified AI model based on the designation of the AI ​​model, and output the plurality of learning data read from the memory to a display device in a predetermined graph form.

[0009] The present disclosure also provides a program for causing a computer having a memory for storing a plurality of pieces of learning data used in training an AI model that predicts the temperature distribution of an object on which components are placed to perform the following process: reading from the memory the plurality of pieces of learning data corresponding to the specified AI model based on the specification of the AI ​​model; and outputting the plurality of pieces of learning data read from the memory to a display device in a predetermined graph format.

[0010] These comprehensive or specific aspects may be realized as a system, an apparatus, a method, an integrated circuit, a computer program, or a recording medium, or may be realized as any combination of a system, an apparatus, a method, an integrated circuit, a computer program, and a recording medium. [Effects of the Invention]

[0011] According to the present disclosure, it is possible to present data used in training an AI surrogate model that predicts the temperature distribution of an object such as a substrate in an intuitive manner. [Brief explanation of the drawings]

[0012] [Figure 1] Conceptual diagram of temperature distribution prediction using AI [Figure 2] FIG. 1 is a block diagram showing an example of a hardware configuration of an information processing device according to an embodiment of the present invention. [Figure 3] A diagram showing an example of the relationship between an AI surrogate model and the training data set used during training. [Figure 4] A graph showing an example of 3D output of training data [Figure 5] A graph showing an example of the first output of the training data in two dimensions [Figure 6] A graph showing an example of the second output of the training data in two dimensions [Figure 7] A flowchart showing an example of the output operation procedure for learning data linked to an AI surrogate model in chronological order. [Figure 8] FIG. 1 is a diagram schematically illustrating an example of a predicted object. [Figure 9]A diagram showing an example of the relationship between the evaluation data and training data sets related to the first update pattern of the AI ​​surrogate model. [Figure 10] A flowchart showing an example of the operation procedure of the first update pattern of the AI ​​surrogate model in chronological order. [Figure 11] A diagram showing an example of the relationship between the evaluation data and training data sets related to the second update pattern of the AI ​​surrogate model. [Figure 12] A flowchart showing an example of the operation procedure for the second update pattern of the AI ​​surrogate model in chronological order. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, with reference to the drawings as appropriate, embodiments specifically disclosing a training data visualization method, a training data visualization device, and a program according to the present disclosure will be described in detail. However, more detailed description than necessary may be omitted. For example, detailed descriptions of well-known matters and redundant descriptions of substantially identical configurations may be omitted. This is to avoid unnecessary redundancy in the following description and to facilitate understanding by those skilled in the art. Note that the accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure, and are not intended to limit the subject matter of the claims.

[0014] First, an overview of predicting temperature distribution using AI will be explained with reference to Figure 1. Figure 1 is a conceptual diagram of predicting temperature distribution using AI.

[0015] Conventionally, the temperature distribution of an object such as a substrate containing a heat-generating element is predicted using simulation techniques such as Computational Fluid Dynamics (CFD) using physical property parameters such as the heat generation amount, thermal conductivity, and emissivity of the heat-generating element. However, predicting the temperature distribution using CFD has the problem that it takes an extremely long time, for example, several hours to several tens of hours, to perform the calculations.

[0016] For example, the target object may be a substrate including multiple elements that are heat-generating bodies. The elements may be, for example, coils, capacitors, diodes, integrated circuits (ICs), or large-scale integration (LSIs). The number, size, and type of elements included in a substrate vary depending on the application, and analyzing the temperature distribution for each substrate using simulation techniques such as CFD would take an enormous amount of time. Note that the target object is not limited to a substrate, and may be any electronic device that includes a heat-generating body, such as an engine or a product that includes a heat-generating body (e.g., a smartphone, a personal computer, etc.).

[0017] Conventional analysis methods take a long time to predict temperature distribution, but by using AI technologies such as Convolutional Neural Networks (CNN), it may be possible to significantly reduce the time required to predict the temperature distribution of an object, from a few seconds to a few tens of seconds.

[0018] Information including physical property parameters such as the heat generation amount of a heating element included in the object, the in-plane thermal conductivity around the heating element, the out-of-plane thermal conductivity and emissivity, etc. When the object is viewed in three dimensions, for example, the input data is information in which physical property parameters are associated with each of the meshes divided into squares by vertical and horizontal mesh division lines in the x-, y-, and z-axis directions.

[0019] By inputting such input data into an AI model, the temperature distribution predicted by the AI ​​model is calculated. Unlike physical models, AI models are generated by machine learning using training data that includes physical property parameters and the predicted temperature distribution. For example, the AI ​​model is a surrogate model (hereinafter referred to as an "AI surrogate model") that uses input data to predict the temperature distribution of an object in which a heating element is located. The physical model here refers to a calculation (simulation technology) that uses physical equations based on classical physics, such as CFD. Using an AI surrogate model to calculate the predicted temperature distribution of an object (hereinafter referred to as the "predicted temperature distribution") can significantly reduce analysis time. Below, an AI surrogate model will be used as an example of an AI model for calculating the temperature distribution of an object.

[0020] However, the use of AI surrogate models is thought to pose the following challenges. Specifically, there is currently no way for anyone other than the creator of the AI ​​surrogate model to know the specific characteristics of the AI ​​surrogate model, such as the type of data (input data) used to train it. For this reason, because the processing of AI surrogate models differs from CFD calculations, which solve predetermined physical equations for the target object to be predicted, it is sometimes difficult for users of AI surrogate models to say that their output is objectively reliable. In other words, users of AI surrogate models need technology that allows them to intuitively understand the data used to train an AI surrogate model that predicts the temperature distribution of an object, such as a circuit board.

[0021] 1. Configuration of the training data visualization device Next, a configuration example of a training data visualization device according to the present disclosure will be described with reference to Fig. 2. Fig. 2 is a block diagram showing an example of a hardware configuration of an information processing device 1 according to the present embodiment. The information processing device 1 is an example of a training data visualization device according to the present disclosure.

[0022] The information processing device 1 learns and stores one or more AI surrogate models that predict the temperature distribution of an object, such as a circuit board on which components including one or more heat-generating elements are arranged. The information processing device 1 is, for example, a computer device such as a desktop personal computer or a stationary server computer. Note that the information processing device 1 is not limited to these computer devices and may be, for example, a portable computer device such as a smartphone or a tablet terminal. The information processing device 1 includes at least a processor 2, a memory 3, an input device 4, and a display device 5. The processor 2, memory 3, input device 4, and display device 5 are connected by wiring such as an internal bus so that data can be input and output to and from each other.

[0023] The processor 2 includes at least one of a central processing unit (CPU), a digital signal processor (DSP), a field programmable gate array (FPGA), and a graphical processing unit (GPU), for example. The processor 2 functions as a controller that manages the overall operation of the information processing device 1. The processor 2 performs control processing for overseeing the operation of each unit of the information processing device 1, data input / output processing between each unit of the information processing device 1, data arithmetic processing, and data storage processing. The processor 2 operates according to a program stored in the memory 3. The processor 2 cooperates with the memory 3 during operation, and temporarily stores data generated or acquired by the processor 2 in the memory 3. In the information processing device 1, the processor 2 cooperates with the memory 3 to predict the temperature distribution of an object by executing an AI surrogate model using data (input data) required to predict the temperature distribution of the object.

[0024] The processor 2 also functions as an information processing device 1, which is an example of a training data visualization device, and includes a training data visualization unit 2a, a training data update unit 2b, and a training unit 2c. The processes performed by the training data visualization unit 2a, the training data update unit 2b, and the training unit 2c are realized based on cooperation between the processor 2 and the memory 3.

[0025] The training data visualization unit 2a reads from the memory 3 a plurality of training data 3a stored therein that correspond to the AI ​​surrogate model 3b specified by the user via the input device 4, and generates a graph with at least two different types of parameters from the plurality of training data 3a as axes. The parameters that make up the training data 3a will be described later with reference to FIG. 3. The training data visualization unit 2a outputs the generated graph to the display device 5 in a predetermined graph format. The graph format displayed on the display device 5 will be described later with reference to FIGS. 4 to 6.

[0026] The learning data update unit 2b updates the learning data 3a corresponding to the AI ​​surrogate model 3b by generating and adding learning data that is insufficient (for example, less than a predetermined number) as the learning data 3a corresponding to the AI ​​surrogate model 3b (see FIG. 10). The learning data update unit 2b also updates the learning data 3a corresponding to the AI ​​surrogate model 3b by deleting learning data that is excessive as the learning data 3a corresponding to the AI ​​surrogate model 3b (see FIG. 12). The addition and deletion of learning data will be described later with reference to FIGS. 10 and 12.

[0027] Each time the learning data update unit 2b updates the number of pieces of learning data 3a, the learning unit 2c performs learning using the updated number of pieces of learning data 3a, updates the corresponding AI surrogate model 3b, and stores it in the memory 3. In this way, by updating an AI surrogate model 3b for which the number of pieces of learning data is insufficient, the learning unit 2c can improve the prediction accuracy of that AI surrogate model 3. Furthermore, by updating an AI surrogate model 3b for which the number of pieces of learning data is surplus, the learning unit 2c can prevent bias in prediction accuracy from occurring due to the influence of the surplus learning data, thereby similarly improving prediction accuracy.

[0028] The memory 3 includes at least a random access memory (RAM) and a read only memory (ROM), for example, and temporarily stores programs necessary for the operation of the information processing device 1 and data acquired or generated during operation. The RAM is, for example, a work memory used during operation of the information processing device 1. The ROM stores and holds, for example, programs for controlling the information processing device 1 in advance. The memory 3 also stores training data 3a and data for an AI surrogate model 3b.

[0029] The training data 3a is a data set that links multiple pieces of input data used in training the corresponding AI surrogate model 3b with prediction results (e.g., predicted temperature distribution) predicted based on the input data. Specific examples of the training data 3a will be described later with reference to FIG. 3.

[0030] The AI ​​surrogate model 3b is an AI surrogate model generated or updated through learning using the corresponding training data 3a, and predicts the temperature distribution of an object, such as a substrate on which one or more heating elements are arranged. The relationship between the training data 3a and the AI ​​surrogate model 3b will be described later with reference to Figure 3. Note that while Figure 3 illustrates the memory 3 storing data for one pair of training data 3a and AI surrogate model 3b, the memory 3 may store data for multiple pairs of training data 3a and AI surrogate model 3b.

[0031] The relationship between training data 3a and AI surrogate model 3b will now be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of the relationship between an AI surrogate model and a data set of training data used during training. As shown in Fig. 3, AI surrogate model 1 is generated by training using training data LM1 configured with a plurality of parameters, and AI surrogate model 2 is generated by training using training data LM2 configured with a plurality of parameters.

[0032] The training data LM1 includes parameters for heat generation (e.g., heat generation density), emissivity, in-plane thermal conductivity, out-of-plane thermal conductivity, grid width (X, Y, Z), and temperature distribution for predicting the temperature distribution of an object, such as a substrate, on which one or more heat generating elements are arranged. The heat generation (e.g., heat generation density), emissivity, in-plane thermal conductivity, out-of-plane thermal conductivity, and grid width (X, Y, Z) correspond to the input data input to the AI ​​surrogate model 1, and the temperature distribution corresponds to the output data predicted by the AI ​​surrogate model 1. The grid width (X, Y, Z) corresponds to the mesh width when the object predicted by the AI ​​surrogate model 1 is divided into three-dimensional (i.e., X-axis, Y-axis, Z-axis) mesh division lines. In other words, the grid width (X, Y, Z) indicates the positions of the heat generating elements arranged on the object, such as a substrate. Although not described further, the training data LM2 has a similar configuration to the training data LM1.

[0033] The input device 4 is a device that receives input from a user of the information processing device 1. The input device 4 may be, for example, a mouse, a keyboard, a touch panel, or a combination of these.

[0034] The display device 5 displays various processing results executed by the information processing device 1. The processing results are, for example, graphs (see FIGS. 4 to 6) visually output by the processor 2 from the learning data 3a corresponding to the AI ​​surrogate model 3b, and prediction results of the temperature distribution of an object such as a substrate predicted by the AI ​​surrogate model 3b. The display device 5 is, for example, a liquid crystal display (LCD) or an organic electroluminescence (EL) display. The input device 4 and the display device 5 may be configured integrally, in which case the input device 4 is a touch panel display or the like.

[0035] 2. Operation of the training data visualization device Next, examples of graph output of the learning data 3a displayed on the display device 5 will be described with reference to Figs. 4 to 6. Fig. 4 is a graph showing a three-dimensional output example of the learning data 3a. Fig. 5 is a graph showing a first two-dimensional output example of the learning data 3a. Fig. 6 is a graph showing a second two-dimensional output example of the learning data 3a.

[0036] The output example of FIG. 4 shows a three-dimensional graph VL1 illustrating the relationship between two parameters (see FIG. 3) of the training data 3a, specifically, the grid width (X, Y) and heat generation density of one training data 3a. The three-dimensional graph VL1 is an example of a visualized graph of training data. That is, the three-dimensional graph VL1 shows the number of training data 3a stored in the memory 3, each plotted with the grid width as the x-axis and the y-axis, and the heat generation density as the z-axis. The three-dimensional graph VL1 is generated by the training data visualization unit 2a of the processor 2 using the training data 3a stored in the memory 3. In the three-dimensional graph VL1, one training data Q1 is represented by the parameters of the grid width (X) corresponding to the x-coordinate, the grid width (Y) corresponding to the y-coordinate, and the heat generation density corresponding to the z-coordinate, and multiple training data are plotted (drawn). This makes it easier for the user to intuitively grasp, by visually checking the three-dimensional graph VL1 output on the display device 5, what kind of training data and how many pieces of training data were used to train the AI ​​surrogate model they selected.

[0037] The output example of FIG. 5 shows a two-dimensional graph VL2 illustrating the relationship between one parameter (specifically, the grid width (X, Y) of one piece of learning data 3a) among multiple parameters (see FIG. 3) of the learning data 3a and the number of pieces of the learning data 3a. The two-dimensional graph VL2 is an example of a visualization graph of learning data. That is, the two-dimensional graph VL2 shows the relationship between the grid width as the x-axis and the y-axis, and the number of pieces of learning data 3a stored in the memory 3. The learning data visualization unit 2a of the processor 2 generates this two-dimensional graph VL2 using the learning data 3a stored in the memory 3. In the two-dimensional graph VL2, one piece of learning data PS1 is represented by the parameters of the grid width (X) corresponding to the x-coordinate and the grid width (Y) corresponding to the y-coordinate, and multiple pieces of learning data are plotted (drawn). This makes it easier for the user to intuitively grasp, by visually checking the two-dimensional graph VL2 output on the display device 5, the position of the heating element on the target object such as a circuit board that was used to train the AI ​​surrogate model that the user selected.

[0038] The output example of FIG. 6 shows a two-dimensional graph VL3 illustrating the relationship between one parameter (specifically, the heat generation density) of one piece of training data 3a among multiple parameters (see FIG. 3) contained in the training data 3a and the number of pieces of training data 3a. The two-dimensional graph VL3 is an example of a visualization graph of training data. That is, the two-dimensional graph VL3 is plotted with the heat generation density on the x-axis and the number of pieces of training data 3a on the y-axis, and the number of pieces of training data 3a is plotted for each piece of training data 3a stored in the memory 3. The training data visualization unit 2a of the processor 2 generates this two-dimensional graph VL3 using the training data 3a stored in the memory 3. The two-dimensional graph VL3 shows a distribution Q2 of the heat generation density of heating elements arranged on an object, such as a substrate, used as the training data 3a, with the number corresponding to the y-coordinate relative to the heat generation density corresponding to the x-coordinate. As a result, by visually checking the two-dimensional graph VL3 output on the display device 5, the user can easily intuitively grasp at what heat density the heat generating elements placed on an object such as a circuit board are evenly distributed in the training data of the AI ​​surrogate model that the user has selected.

[0039] Next, the output operation procedure of the learning data 3a according to this embodiment will be described with reference to Fig. 7. Fig. 7 is a flowchart showing an example of the output operation procedure of learning data linked to an AI surrogate model in chronological order. The series of processes shown in Fig. 7 are realized by the processor 2 of the information processing device 1 working in cooperation with the memory 3.

[0040] 7, processor 2 receives, through a user operation, from input device 4, the selection of an AI surrogate model for which training data is to be output (step St1). Processor 2 reads and acquires the training data used during training of the AI ​​surrogate model selected in step St1 from memory 3. Processor 2 generates a display screen (not shown) that displays multiple parameters (e.g., grid width (X, Y, Z), heat generation density, etc.) contained in the read training data, and outputs the display screen to display device 5 (step St2).

[0041] The processor 2 selects parameters for the visualized graph of the learning data specified by a user operation from the input device 4 from among the items on the display screen displayed in step St2 (step St3). The processor 2 receives a selection of the display format of the visualized graph of the learning data from the input device 4 by a user operation (step St4). The display format may be, for example, the two-dimensional graph VL2 in FIG. 5 or the two-dimensional graph VL3 in FIG. 6, or the three-dimensional graph VL1 in FIG. 4. Note that the processing of step St4 is not limited to being received from the input device 4, and may also be executed according to instructions for the display format written in a setting file (not shown) that is generated in advance and stored in the memory 3.

[0042] Processor 2 generates a visualized graph of the learning data based on the parameters selected in step St3, in accordance with the display format selected in step St4, and outputs the visualized graph to display device 5 (step St5). By visually checking the visualized graph of the learning data output to display device 5, the user can intuitively grasp what kind of learning data was used to train the AI ​​surrogate model they selected, and can clearly understand the characteristics of the AI ​​surrogate model.

[0043] The visualized graphs of the training data shown in Figures 4 to 6 allow the user to intuitively grasp the training data used in training the AI ​​surrogate model. At the same time, the visualized graphs also allow the user to grasp whether there is insufficient training data or whether there is excessive duplication of training data. In other words, updating the AI ​​surrogate model by adding insufficient training data or deleting excessive training data is expected to further improve the prediction accuracy of the AI ​​surrogate model.

[0044] Assuming that a visualized graph of the training data has been displayed, the following describes the process of updating the AI ​​surrogate model by adding insufficient training data or deleting excess training data, with reference to FIGS. 8 to 12.

[0045] First, with reference to Figs. 8 to 10, an overview of the process of updating an AI surrogate model by adding insufficient training data will be described. Fig. 8 is a diagram schematically illustrating an example of an object to be predicted. Fig. 9 is a diagram illustrating an example of the relationship between evaluation data and training data sets related to a first update pattern of the AI ​​surrogate model. Fig. 10 is a flowchart illustrating an example of the operation procedure of the first update pattern of the AI ​​surrogate model 3b in chronological order. The series of processes shown in Fig. 10 are implemented by the processor 2 of the information processing device 1 working in cooperation with the memory 3.

[0046] FIG. 8 shows the substrate SUB1, which is the object to be predicted, in a plan view along two axes, the x-axis and the y-axis. Two heating elements HE1 and HE2 are arranged on this substrate SUB1. Here, it is assumed that the temperature distribution of the substrate SUB1 shown in FIG. 8 is predicted using the AI ​​surrogate model 3b already stored in the memory 3 as evaluation data. Before performing this prediction, the process of updating the AI ​​surrogate model 3b by adding missing learning data to the current learning data 3a will be described with reference to FIGS. 9 and 10.

[0047] Fig. 9 shows a two-dimensional graph VL2a, which is a visualized graph of the training data 3a similar to Fig. 5. That is, the number of pieces of training data 3a plotted for each heating element position along two axes, the x-axis and the y-axis, is shown, corresponding to the number of pieces of training data 3a already stored in the memory 3. However, for convenience of explanation, Fig. 9 additionally plots training data HE2U, HE2L, and HE2R that would be added around the heating element HE2 (described later) to the two-dimensional graph VL2 of Fig. 5.

[0048] In FIG. 10, the processor 2 determines, based on the learning data 3a stored in the memory 3, whether or not existing learning data 3a exists in the 10 squares surrounding each of the heating elements HE1 and HE2 in the evaluation data shown in FIG. 9 (step St11). Here, one square corresponds to one grid (i.e., a mesh divided by mesh dividing lines). Furthermore, the 10 squares are merely an example and are not limited to 10 squares. For simplicity of explanation, an example of determining whether or not existing learning data 3a exists in the 10 squares surrounding the heating element HE2 will be described here, but the following explanation can be similarly applied to the explanation of an example of determining whether or not existing learning data 3a exists in the 10 squares surrounding the heating element HE1.

[0049] If the processor 2 determines that no existing learning data 3a exists at the upper boundary of the 10-square virtual circle IGC1 around the heating element HE2 (if it is not above the heating element), it decides to place the learning data HE2U at the upper boundary of the IC (i.e., the heating element) (step St12a). Furthermore, the processor 2 sets a heat density for the learning data HE2U placed in step St12a such that the number of existing learning data 3a is small (for example, less than a predetermined value) (step St13a). This enables the processor 2 to add the learning data HE2U to the upper boundary of the 10-square virtual circle IGC1 around the heating element HE2 in the evaluation data, thereby increasing the number of existing learning data 3a stored in the memory 3.

[0050] If the processor 2 determines that no existing learning data 3a exists at the right boundary position of the imaginary circle IGC1 of 10 squares around the heating element HE2 (if it is not to the right), it decides to place the learning data HE2R at the right boundary position of the IC (i.e., the heating element) (step St12b). Furthermore, the processor 2 sets a heat density for the learning data HE2R placed in step St12b such that the number of existing learning data 3a is small (for example, less than a predetermined value) (step St13b). This allows the processor 2 to add the learning data HE2R to the right boundary position of the imaginary circle IGC1 of 10 squares around the heating element HE2 in the evaluation data, thereby increasing the number of existing learning data 3a stored in the memory 3.

[0051] If the processor 2 determines that no existing learning data 3a exists at the left boundary position of the imaginary circle IGC1 of 10 squares around the heating element HE2 (if it is not to the left), it decides to place the learning data HE2L at the left boundary position of the IC (i.e., the heating element) (step St12c). Furthermore, the processor 2 sets a heat density in which the number of existing learning data 3a is small (for example, less than a predetermined value) to the learning data HE2L placed in step St12c (step St13c). This allows the processor 2 to add the learning data HE2L to the left boundary position of the imaginary circle IGC1 of 10 squares around the heating element HE2 in the evaluation data, thereby increasing the number of existing learning data 3a stored in the memory 3.

[0052] If processor 2 determines that no existing learning data 3a exists at the lower boundary of the 10-square virtual circle surrounding the heating element (if it is not below the heating element), it decides to place learning data at the lower boundary of the IC (i.e., the heating element) (step St12d). Furthermore, processor 2 sets a heat density for the learning data placed in step St12d that results in a small number of existing learning data 3a (e.g., less than a predetermined value) (step St13d). This allows processor 2 to add learning data to the lower boundary of the 10-square virtual circle surrounding the heating element in the evaluation data, thereby appropriately increasing the number of existing learning data 3a stored in memory 3. Note that in the example of FIG. 9, the lower boundary of the 10-square virtual circle IGC1 surrounding the heating element HE2 exceeds the area of ​​the substrate SUB1, so adding learning data at that lower boundary is not possible.

[0053] On the other hand, if the processor 2 determines that existing learning data 3a exists at any of the boundary positions of the 10-square virtual circle IGC1 around the heating element HE2 (if there is any), the processor 2 generates and references the heat generation distribution density (see FIG. 6) of the learning data 3a based on the existing learning data 3a stored in the memory 3 (step St14). If the processor 2 determines that there is no range in the heat generation density distribution generated in step St14 where no heat generation distribution data exists (if there is any data), the processing of the processor 2 ends.

[0054] On the other hand, if the processor 2 determines that there is a range in the heat density distribution generated in step St14 in which no data exists (if there is a range in which there is no data), it sets the heat density in the range in which there is no data to be added as learning data in the evaluation data (see FIG. 8) (step St15). For example, the data of the heat density "0.08" in FIG. 6 in which there is no data is added.

[0055] After processing at least one of steps St13a to St13d or after processing step St15, the processor 2 performs a CFD calculation based on the current learning data, including the added learning data, to calculate a predicted temperature distribution of the evaluation data (see FIG. 8) (step St16). The processor 2 adds the calculation result of step St16 (i.e., the predicted temperature distribution of the substrate SUB1 corresponding to the evaluation data) as learning data in the evaluation data (see FIG. 8) (step St17). The processor 2 uses the learning data 3a added in step St17 to train the corresponding AI surrogate model 3b (step St18) and updates the AI ​​surrogate model 3b (step St19). In this way, the processor 2 updates the AI ​​surrogate model 3b by adding (increasing) learning data with a small number of learning data or in which some parameters (e.g., heat generation density) are not present, thereby further improving the prediction accuracy of the AI ​​surrogate model 3b.

[0056] Next, an overview of the process of deleting excess training data and updating the AI ​​surrogate model will be described with reference to FIGS. 11 and 12. FIG. 11 is a diagram showing an example of the relationship between the evaluation data and the training data datasets associated with the second update pattern of the AI ​​surrogate model. FIG. 12 is a flowchart showing a chronological example of the operation procedure of the second update pattern of the AI ​​surrogate model 3b. The series of processes shown in FIG. 12 are implemented by the processor 2 of the information processing device 1 working in cooperation with the memory 3. In the explanation of FIGS. 11 and 12, it is assumed that the evaluation data, which is the object to be predicted, is the substrate SUB1 of FIG. 8, and that the heat generation density distribution, which is a visualized graph of the training data 3a stored in the memory 3, is the same as that shown in the two-dimensional graph VL3 of FIG. 6.

[0057] Fig. 11 shows a two-dimensional graph VL2a, which is a visualized graph of the learning data 3a similar to Fig. 5. That is, the number of pieces of learning data 3a already stored in the memory 3, plotted for each position of the heating element along two axes, the x-axis and the y-axis, is shown.

[0058] 12, the processor 2 determines whether or not there is any learning data 3a in which the positions of heating elements overlap in the two-dimensional graph VL2b showing the distribution of the positions of the heating elements shown in Fig. 11, based on the learning data 3a stored in the memory 3 (step St21). If it is determined that there is no learning data 3a in which the positions of heating elements overlap (no), the processing of the processor 2 proceeds to step St25.

[0059] On the other hand, if processor 2 determines that there is learning data 3a in which the positions of heat generating elements overlap (see overlapping portion IGC2 in FIG. 11), processor 2 checks (sees) the heat generation density of the heat generating element (e.g., heat generating element HE3 in FIG. 11) closer to the center position ICC of substrate SUB1 (see FIG. 8) (step St22). Specifically, processor 2 determines whether the heat generation density of the heat generating element (e.g., heat generating element HE3 in FIG. 11) closer to the center position ICC is equal to the values ​​at both ends (i.e., the maximum and minimum values) of the heat generation density distribution (see FIG. 6).

[0060] If the processor 2 determines (step St23) that the heat density of the heat generating element closer to the center position ICC (for example, the heat generating element HE3 in FIG. 11) is not equal to the values ​​at both ends of the heat generating density distribution (see FIG. 6), it deletes the learning data corresponding to that heat generating density. In other words, if there are already multiple pieces of learning data with overlapping positions, the learning data closest to the center position ICC of the substrate SUB1 is considered to be redundant learning data. Based on this determination, the processor 2 deletes that learning data, thereby reducing the number of learning data 3a to be stored in the memory 3.

[0061] On the other hand, if the processor 2 determines (step St24) that the heat density of the heat generating element closer to the center position ICC (for example, the heat generating element HE3 in FIG. 11) is equal to the values ​​at both ends of the heat generating density distribution (see FIG. 6), it retains the learning data corresponding to that heat generating density without deleting it. In other words, even if there are already multiple pieces of learning data with overlapping positions, the learning data closest to the center position ICC of the substrate SUB1 is considered to be learning data that does not exist in excess so as not to predict that the heat generating density deviates from the boundary value (i.e., the maximum or minimum value) of the heat generating density distribution. Based on this determination, the processor 2 retains that learning data, thereby maintaining the number of pieces of learning data 3a to be stored in the memory 3.

[0062] When the processor 2 determines that there is no training data 3a in which the positions of heat generating elements overlap or that there are no longer any training data 3a (if there are no training data), the processor 2 trains the corresponding AI surrogate model 3b using the number of training data 3a currently stored in the memory 3 (step St25) and updates the AI ​​surrogate model 3b (step St26). In this way, the processor 2 determines whether or not adjustment (i.e., deletion) is required for the number of surplus training data based on whether or not the number corresponds to a boundary value of the heat density distribution, and updates the AI ​​surrogate model 3b based on an appropriate number of training data 3a, thereby further improving the prediction accuracy of the AI ​​surrogate model 3b.

[0063] Summary of the Disclosure The above description of the embodiments discloses technical concepts corresponding to the following items.

[0064] (Item 1) The learning data visualization method according to the present disclosure includes storing, in a memory (3), a plurality of learning data sets used in learning an AI model (e.g., an AI surrogate model) that predicts the temperature distribution of an object (e.g., a circuit board) on which a component (e.g., a heat generating element) is arranged; Based on the designation of the AI ​​model, the plurality of learning data corresponding to the designated AI model are read from the memory; The plurality of learning data read from the memory are output to a display device (5) in a predetermined graph format. As a result, the learning data visualization method makes it possible to present data used in learning an AI surrogate model that predicts the temperature distribution of an object such as a substrate in an intuitive manner.

[0065] (Item 2) In the training data visualization method according to item 1, the training data is configured to have a plurality of types of parameters, The graph format outputs a graph with axes of at least two different types of parameters among the plurality of types of parameters. As a result, the training data visualization method can present the training data to the user in an intuitive and easy-to-understand manner as a graph with axes representing multiple different types of parameters that make up the training data.

[0066] (Item 3) In the training data visualization method according to item 1 or 2, the plurality of types of parameters include at least a position of a component and a heat density of the component. As a result, the learning data visualization method can specifically present, as examples of parameters, a distribution example showing the location of learning data for each component position, and a heat density distribution example showing the proportion of component heat density values.

[0067] (Item 4) In the learning data visualization method described in item 3, a new object (e.g., a substrate SUB1) on which components (e.g., heating elements HE1 and HE2) are arranged is prepared as evaluation data, If there is no learning data having a parameter of a position within a predetermined range from the component of the evaluation data (for example, a position within 10 squares around the component) among the plurality of learning data stored in the memory, a component is newly placed at a position within the predetermined range, and a heat density is set such that the number of heat density data included in the plurality of learning data is less than a predetermined value; calculating a temperature distribution corresponding to the evaluation data based on the newly placed positions of the components and the set heat generation density; New learning data including the positions of the newly placed components, the set heat density, and the calculation results of the temperature distribution are added. As a result, the learning data visualization method makes it possible to add learning data to positions within a specified range of the heating element in the evaluation data (for example, the upper, left, right, and lower boundary positions of a virtual circle with 10 squares around it), making it possible to appropriately increase the number of existing learning data stored in memory.

[0068] (Item 5) In the training data visualization method described in Item 4, after the new training data is added, the specified AI model is updated using the new training data that has been added and the plurality of training data stored in the memory. As a result, the training data visualization method can further improve the prediction accuracy of the AI ​​surrogate model by updating the AI ​​surrogate model by adding (increasing) training data that has a small number of training data or that does not contain some parameters (e.g., heat density).

[0069] (Item 6) In the training data visualization method described in Item 3, if there are two or more pieces of training data in which the positions of parts overlap among the plurality of pieces of training data stored in the memory, and the heat density of a part close to the center position of the object corresponding to the training data is neither the maximum value nor the minimum value of the heat density included in the plurality of training data, the training data corresponding to that part is deleted from the plurality of training data stored in the memory. As a result, the training data visualization method makes it possible to appropriately determine whether or not there is surplus training data that is not actually needed, and to appropriately reduce the amount of training data that needs to be stored in memory in order to ensure the predictive accuracy of the AI ​​surrogate model.

[0070] (Item 7) In the training data visualization method described in item 6, the specified AI model is updated using the plurality of training data stored in the memory after the training data has been deleted. As a result, according to the training data visualization method, the need for adjustment (i.e., deletion) of surplus training data is determined based on whether the number of pieces of training data corresponds to a boundary value of the heat density distribution, and the AI ​​surrogate model can be updated based on an appropriate number of training data, thereby further improving the prediction accuracy of the AI ​​surrogate model.

[0071] (Item 8) The training data visualization device according to the present disclosure includes: a processor; a memory that stores a plurality of pieces of learning data used in learning an AI model that predicts the temperature distribution of an object on which components are arranged; The processor, in cooperation with the memory, Based on the designation of the AI ​​model, the plurality of learning data corresponding to the designated AI model are read from the memory; The plurality of learning data read from the memory is output to a display device in a predetermined graph format. As a result, the learning data visualization device can present data used in learning an AI surrogate model that predicts the temperature distribution of an object such as a substrate in an intuitive manner.

[0072] (Item 9) The program according to the present disclosure is A computer having a memory for storing a plurality of pieces of learning data used in learning an AI model that predicts the temperature distribution of an object on which a component is placed, a process of reading, from the memory, the plurality of pieces of learning data corresponding to the designated AI model based on the designation of the AI ​​model; and outputting the plurality of learning data read from the memory to a display device in a predetermined graph format. This allows the computer to present data used to train an AI surrogate model that predicts the temperature distribution of an object such as a substrate in an intuitive manner.

[0073] Although the embodiments have been described above with reference to the accompanying drawings, the present disclosure is not limited to such examples. It is clear that a person skilled in the art can conceive of various modifications, alterations, substitutions, additions, deletions, and equivalents within the scope of the claims, and it is understood that these also fall within the technical scope of the present disclosure. Furthermore, the components in the above-described embodiments may be combined in any manner without departing from the spirit of the invention. [Industrial Applicability]

[0074] The technology disclosed herein is useful as a training data visualization method, training data visualization device, and program that present data used in training an AI surrogate model that predicts the temperature distribution of an object such as a substrate in an intuitive manner. [Explanation of symbols]

[0075] 1. Information processing equipment 2 processors 2a Learning data visualization part 2b Learning data update section 2c Learning Department 3. Memory 4 Input Devices 5 Display Devices

Claims

1. storing in a memory a plurality of pieces of learning data used in learning an AI model that predicts the temperature distribution of an object on which components are arranged; Based on the designation of the AI ​​model, reading the plurality of learning data corresponding to the designated AI model from the memory; outputting the plurality of learning data read from the memory to a display device in a predetermined graph format; Training data visualization methods.

2. the learning data includes a plurality of types of parameters; The graph format outputs a graph with axes of at least two different types of parameters among the plurality of types of parameters. The training data visualization method according to claim 1 .

3. the plurality of types of parameters include at least a position of a component and a heat density of the component; The training data visualization method according to claim 2 .

4. A new object with parts placed on it is prepared as evaluation data. when there is no learning data having a parameter of a position within a predetermined range from the component of the evaluation data among the plurality of learning data stored in the memory, a component is newly placed at a position within the predetermined range, and a heat density is set such that the number of heat density data included in the plurality of learning data is less than a predetermined value; calculating a temperature distribution corresponding to the evaluation data based on the newly placed positions of the components and the set heat generation density; adding new learning data including the positions of the newly placed components, the set heat generation density, and the calculation results of the temperature distribution; The training data visualization method according to claim 3 .

5. After the new learning data is added, the specified AI model is updated using the new learning data that has been added and the plurality of learning data stored in the memory. The training data visualization method according to claim 4 .

6. If there are two or more pieces of learning data in which the positions of components overlap among the plurality of pieces of learning data stored in the memory, and the heat density of a component close to the center position of the object corresponding to the learning data is neither the maximum value nor the minimum value of the heat density included in the plurality of learning data, delete the learning data corresponding to that component from the plurality of learning data stored in the memory. The training data visualization method according to claim 3 .

7. updating the specified AI model using the plurality of learning data stored in the memory after the learning data has been deleted; The training data visualization method according to claim 6 .

8. a processor; a memory that stores a plurality of pieces of learning data used in learning an AI model that predicts the temperature distribution of an object on which components are arranged, The processor, in cooperation with the memory, Based on the designation of the AI ​​model, reading the plurality of learning data corresponding to the designated AI model from the memory; outputting the plurality of learning data read from the memory to a display device in a predetermined graph format; A training data visualization device.

9. A computer having a memory for storing a plurality of pieces of learning data used in learning an AI model that predicts the temperature distribution of an object on which a component is arranged, A process of reading out the plurality of learning data corresponding to the designated AI model from the memory based on the designation of the AI ​​model; and outputting the plurality of learning data read from the memory to a display device in a predetermined graph format. program.

Images