Ceramic product design support method and ceramic product design support system

Abstract

To evaluate a honeycomb product with high accuracy in a short time in designing the honeycomb product.SOLUTION: The system learns a machine learning model that substitutes for a physical simulation in which a product condition of a honeycomb product is input and a product characteristic of the honeycomb product is output, using a data set including a set of the product condition input to the physical simulation and the product characteristic output from the physical simulation as a result of inputting the product condition to the physical simulation. The system predicts product characteristics of a target product, which is a honeycomb product to be designed, by inputting product conditions of the target product to the learned machine learning model, and outputs evaluation result information, which is information representing the predicted product characteristics of the target product.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates generally to design support for ceramic products, and more particularly to design support for ceramic products having honeycomb structures. [Background technology]

[0002] A technology relating to support for designing ceramic products is disclosed in Patent Document 1. According to the technology disclosed in Patent Document 1, ceramic products are designed using simulation software. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-059645 Summary of the Invention [Problem to be solved by the invention]

[0004] Ceramic products include those with honeycomb structures (hereinafter referred to as honeycomb products). Honeycomb products can be used in a variety of fields, such as air purification (e.g., automobile catalysts), gas compression, and heat exchange.

[0005] The product number of a honeycomb product varies depending on the type of device (for example, vehicle model) in which the honeycomb product is installed, the customer who purchases the honeycomb product, the destination of the honeycomb product, etc. For this reason, there are many product numbers for honeycomb products.

[0006] Honeycomb products are designed for each product number. When designing honeycomb products for each product number, the honeycomb products are evaluated (typically, the performance or reliability of the honeycomb products is evaluated). Methods for evaluating honeycomb products include experiments and physical simulations, but experiments alone cannot reveal the internal behavior of a honeycomb product (for example, the continuous or local temperature distribution of the ribs that make up the honeycomb product, or the stress values ​​of the ribs). For this reason, detailed evaluation using physical simulations is required.

[0007] Physical simulations generally involve discretizing and solving the governing partial differential equations, which takes time. For example, a thermal stress analysis of a honeycomb product requires an analysis lead time of the order of weeks (e.g., one to two weeks). Because many honeycomb product models are designed each year, it is desirable to shorten the time required to evaluate honeycomb products. However, there is a trade-off between the accuracy of physical simulations and the calculation costs (e.g., calculation load and calculation time), and reducing the calculation costs to save time results in a decrease in the accuracy of the physical simulation. [Means for solving the problem]

[0008] The system trains a machine learning model that replaces a physical simulation that inputs product conditions of a honeycomb product and outputs product characteristics of the honeycomb product, using a data set that includes pairs of the product conditions input to the physical simulation and the product characteristics output from the physical simulation as a result of inputting the product conditions into the physical simulation.By inputting the product conditions of a target product, which is a honeycomb product to be designed, into the trained machine learning model, the system predicts the product characteristics of the target product and outputs evaluation result information that is information representing the predicted product characteristics of the target product. [Effects of the Invention]

[0009] According to the present invention, evaluation of a honeycomb product can be performed with high accuracy and in a short time when designing the honeycomb product. [Brief explanation of the drawings]

[0010] [Figure 1] 1 shows an example of the configuration of a ceramic product design support system according to an embodiment of the present invention. [Figure 2] 1 shows an example of a flow of a thermal stress analysis process according to an embodiment. [Figure 3] 10 shows an example of a flow of a thermal stress analysis process according to a comparative example. [Figure 4] An example of preparing a training dataset is shown below. [Figure 5] An example of model creation (learning) is shown below. [Figure 6] An example of model utilization (inference) is shown below. [Figure 7] An example of a training dataset is shown below. [Figure 8] An example of the accuracy verification result is shown below. DETAILED DESCRIPTION OF THE INVENTION

[0011] In the following description, an "interface apparatus" may refer to one or more interface devices. The one or more interface devices may be at least one of the following: One or more I / O (Input / Output) interface devices. The I / O (Input / Output) interface devices are interface devices to at least one of the I / O device and a remote display computer. The I / O interface device to the display computer may be a communications interface device. The at least one I / O device may be a user interface device, for example, either an input device such as a keyboard and a pointing device, or an output device such as a display device. One or more communication interface devices. The one or more communication interface devices may be one or more homogeneous communication interface devices (e.g., one or more NICs (Network Interface Cards)) or two or more heterogeneous communication interface devices (e.g., an NIC and an HBA (Host Bus Adapter)).

[0012] In the following description, "memory" refers to one or more memory devices, which are an example of one or more storage devices, and may typically be a primary storage device. At least one memory device in the memory may be a volatile memory device or a non-volatile memory device.

[0013] In the following description, a "persistent storage device" may refer to one or more persistent storage devices, which are an example of one or more storage devices. A persistent storage device may typically be a non-volatile storage device (e.g., an auxiliary storage device), and more specifically, may be, for example, a hard disk drive (HDD), a solid state drive (SSD), a non-volatile memory express (NVME) drive, or a storage class memory (SCM).

[0014] In the following description, the term "storage device" may refer to at least one of memory and persistent storage device.

[0015] Furthermore, in the following description, a "processor" may refer to one or more processor devices. The at least one processor device may typically be a microprocessor device such as a CPU (Central Processing Unit), but may also be another type of processor device such as a GPU (Graphics Processing Unit). The at least one processor device may be a single-core or multi-core. The at least one processor device may also be a processor core. The at least one processor device may also be a processor device in a broader sense, such as a circuit that is a collection of gate arrays written in a hardware description language that performs part or all of the processing (for example, an FPGA (Field-Programmable Gate Array), a CPLD (Complex Programmable Logic Device), or an ASIC (Application Specific Integrated Circuit)).

[0016] Furthermore, in the following description, functions may be described using the expression "yyy unit." However, the functions may be realized by one or more computer programs executed by a processor, by one or more hardware circuits (e.g., FPGAs or ASICs), or by a combination thereof. When a function is realized by a program executed by a processor, the specified processing is performed using a storage device and / or an interface device, etc., as appropriate, and therefore the function may be considered to be at least a part of the processor. Processing described using a function as the subject may be processing performed by a processor or a device having the processor. A program may be installed from a program source. The program source may be, for example, a computer from which the program is distributed or a computer-readable recording medium (e.g., a non-transitory recording medium). The description of each function is an example, and multiple functions may be combined into one function, or one function may be divided into multiple functions.

[0017] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0018] FIG. 1 shows an example of the configuration of a ceramic product design support system according to one embodiment of the present invention.

[0019] A ceramic product design support system 50 is constructed as a data processing system that supports the design of honeycomb products. The ceramic product design support system 50 may be a physical computer system, a logical computer system based on a physical computer system, or a combination of at least a part of a physical computer system and at least a part of a logical computer system. The physical computer system may be composed of one or more physical computers and may include an interface device, a storage device, and a processor connected to the physical computers. The logical computer system may include a virtual machine or a system as a cloud computing service.

[0020] In this embodiment, the ceramic product design support system 50 includes one or more edge systems 100 and a core system 150 that communicates with the one or more edge systems 100 via a communication network 70 (e.g., the Internet or a Wide Area Network (WAN)). Each of the edge system 100 and the core system 150 is a physical computer system in this embodiment, but may alternatively or additionally be a logical computer system.

[0021] The edge system 100 includes an interface device 101, a storage device 102, and a processor 103 connected thereto. The edge system 100 may be a personal computer or a server computer that communicates with a client computer as an input / output console having an input device 104 and a display device 105.

[0022] An input device 104 (e.g., a keyboard and a pointing device) and a display device 105 are connected to the interface device 101. The input device 104 and the display device 105 may be, for example, a touch panel. In addition, a core system 150 is communicatively connected to the interface device 101 via a communication network 70.

[0023] The storage device 102 stores data and programs. For example, the storage device 102 stores one or more simulation models 121 and one or more surrogate models 122. Each simulation model 121 may be a mathematical model and typically includes a partial differential equation. A simulation unit 131 (described later) converts the partial differential equation in the simulation model 121 into a discretized algebraic equation and solves the algebraic equation. At least one simulation model 121 may include an algebraic equation resulting from discretizing the partial differential equation. On the other hand, the surrogate model 122 is a model that replaces a physical simulation and is a machine learning model 171 (i.e., a trained machine learning model 171) that has been trained and deployed to the edge system 100.

[0024] Although the simulation model 121 and the surrogate model 122 may correspond to each other on a one-to-one basis, it is preferable that they are prepared in a one-to-many (or many-to-many) ratio.

[0025] Specifically, the simulation model 121 may be prepared for each function (for each type of physical quantity to be calculated). For example, a simulation model 121 as a temperature calculation model and a simulation model 121 as a stress calculation model may be prepared.

[0026] One or more surrogate models 122 are prepared for each simulation model 121. Typically, it is preferable to prepare multiple surrogate models 122 for one simulation model 121, each trained on data under different conditions. This is because if one machine learning model 171 is trained to cover all honeycomb product part numbers, the surrogate model 122 serving as the machine learning model 171 after the training may be highly versatile but may have poor accuracy for specific product numbers. For example, even if the physical simulation (simulation model 121) is the same for Cd honeycomb products and SiC honeycomb products, it is preferable to train the machine learning model 171 separately. As a result, two different surrogate models 122, such as a surrogate model 122 for Cd honeycomb products and a surrogate model 122 for SiC honeycomb products, are prepared. Furthermore, by training the machine learning model 171 by narrowing the part numbers and conditions to a certain range, it is possible to prepare a surrogate model 122 with high accuracy within that range. Furthermore, surrogate models 122 may be prepared for each learning method (for example, for each model type, such as a linear regression model or a neural network model) for the same simulation model 121. In this way, for each simulation model 121, a number of surrogate models 122 may be prepared that is determined according to at least one of the requirements for prediction accuracy, the amount of teacher data set to be prepared, the complexity of the calculation (simulation model 121), and the range of the learning target (product number, conditions, etc.).

[0027] When the processor 103 executes a computer program stored in the storage device 102, functions such as a simulation unit 131, an inference unit 132, and an output unit 133 are realized. The simulation unit 131 executes a physical simulation (e.g., finite element analysis) using a simulation model 121 (e.g., a model based on the finite element method). The inference unit 132 performs inference using a surrogate model 122. The output unit 133 outputs information based on the inference result. The output of the information based on the inference result may be by displaying the information on the display device 105 or by transmitting the information to the core system 150 (or another computer system).

[0028] The core system 150 includes an interface device 151, a storage device 152, and a processor 153 connected thereto. The edge system 150 may be a server computer that communicates with the edge system 100.

[0029] The interface device 151 is communicatively connected to the edge system 100 via the communication network 70 .

[0030] The storage device 152 stores data and programs. For example, the storage device 152 stores one or more machine learning models 171 and a training dataset 172 for each of the one or more machine learning models 171.

[0031] The computer program stored in the storage device 152 is executed by the processor 153 to realize functions such as a learning unit 181. The learning unit 181 learns the machine model 171 using the teacher dataset 172 of the machine learning model 171.

[0032] Since physical simulations require a high computational load, the simulation model 121 and the simulation unit 131 may be provided in the core system 150, which can have more computational resources than the edge system 100.

[0033] Below, the details will be explained using thermal stress analysis as an example of honeycomb product evaluation.

[0034] FIG. 2 shows an example of the flow of a thermal stress analysis process according to the embodiment.

[0035] For example, a thermal stress analysis process is performed for each honeycomb product model number that requires a thermal stress analysis. The thermal stress analysis process is an example of a honeycomb product evaluation process and is a part of a honeycomb product design process.

[0036] The thermal stress analysis process according to the embodiment can be composed of only a meeting process and a document preparation process. As will be described in detail later, this embodiment uses a surrogate model 122, a high-speed calculation model that simulates thermal stress analysis, enabling thermal stress analysis in a shorter time (e.g., real time) than physical simulation. Therefore, during a face-to-face or online meeting (i.e., during the meeting process) between a honeycomb product analyst (an example of an evaluator) 201 and a honeycomb product designer 202, they can coordinate requirements for the honeycomb product of the target product number, perform real-time thermal stress analysis using the surrogate model 122, share analysis results between the analyst 201 and the designer 202 (and, if necessary, coordinate requirements again, as indicated by the dashed arrows), and discuss the flow of report materials. After the meeting process, in the document preparation process, the analyst 201 prepares report materials for the designer 202.

[0037] As a result, the time required for the thermal stress analysis process according to the embodiment (analysis lead time) is, for example, one day.

[0038] In the consultation process, the "requests for the honeycomb product" include the product conditions of the honeycomb product. The product conditions may include at least one of conditions related to the structure of the honeycomb product, operating conditions of the honeycomb product (e.g., the level (e.g., value or value range) for each operating condition item), and conditions inside the honeycomb product. When the evaluation is a thermal stress analysis, the product conditions include the structure of the honeycomb product and the temperature distribution inside the honeycomb product. As a result, the output of the thermal stress analysis using the surrogate model 122 includes the stress distribution inside the honeycomb product as a product characteristic.

[0039] In the document creation process, the analyst 201 may perform a supplementary thermal stress analysis, for example, by inputting product conditions not used in the consultation process into the surrogate model 122 to obtain product characteristics under those product conditions and describe them in the report. The report is provided to the designer 202. "Providing the report" may mean sending paper or digital report materials by mail or email, or may mean inputting the report materials into the ceramic product design support system 50 or another computer system and displaying them on the information processing terminal of the designer 202 via that system. In this way, the report is shared between the analyst 201 and the designer 202.

[0040] FIG. 3 shows an example of the flow of a thermal stress analysis process according to a comparative example.

[0041] In the comparative example, the thermal stress analysis is performed by physical simulation without the surrogate model 122. This makes the analysis time-consuming, and it is difficult to share the analysis results between the analyst 201 and the designer 202 during the meeting process. Therefore, the thermal stress analysis process according to the comparative example has a preparation process, an analysis process, and a confirmation process between the meeting process and the document creation process, and a sharing process after the document creation process.

[0042] In the meeting process, only the reconciliation of requirements takes place. In the preparation process, the analyst 201 determines the product conditions to be input into the physical simulation based on the results of the reconciliation of requirements. In the analysis process, the physical simulation is executed. In the confirmation process, the analyst 201 confirms the results of the physical simulation. In the document creation process, the analyst 201 creates a report based on the confirmed physical simulation results. In the sharing process, the report created by the analyst 201 is shared with the designer 202.

[0043] As a result, the time required for the thermal stress analysis process (analysis lead time) for the comparative example is 1 to 2 weeks. If the report does not match the design intent, the product conditions must be changed and recalculated (repeated physical simulation), and the preparation process must be carried out again. If this repetition occurs, the analysis lead time becomes even longer.

[0044] Hereinafter, the honeycomb product design support according to the embodiment will be described in detail, taking thermal stress analysis as an example of honeycomb product evaluation, similarly to Figs.

[0045] The honeycomb product design support method according to the embodiment includes preparation of a training data set, model creation (learning), and model utilization (inference).

[0046] FIG. 4 shows an example of preparing the training data set 172.

[0047] The training data set 172 is a set of training data used in model creation (learning). The training data includes a set of input data and output data corresponding to the input data. The input data includes data related to product conditions, and the output data includes data related to product characteristics.

[0048] The input data may be prepared by experiment or by simulation using CAE (Computer Aided Engineering) or the like. Here, temperature distribution data is prepared for each product number. Specifically, for example, a predetermined number of time series of temperature distribution data (for example, a predetermined pattern) are prepared for each product number.

[0049] The output data is data output through a physical simulation of thermal stress analysis, and is stress distribution data. Specifically, for example, for each piece of time-series temperature distribution data, a physical simulation is executed using the temperature distribution data as input, and the output data obtained by executing the physical simulation is stress distribution data. As a result, a time-series of stress distribution data corresponding to the time-series of temperature distribution data is prepared through the physical simulation. The physical simulation is executed by the simulation unit 131 using the simulation model 121 for thermal stress analysis.

[0050] In this way, the teacher data in the teacher data set 172 includes a set of input data prepared by experiments, CAE, or the like, and output data obtained by a physical simulation using the input data. A teacher data set 172 is prepared for each machine learning model 171. To create (construct) a surrogate model 122 that represents a physical simulation, the machine learning model 171 is trained, and a teacher data set 172 is prepared for this training. The output data in the teacher data is data obtained through a physical simulation using a simulation model 121 that corresponds to the machine learning model 171. In this way, a surrogate model 122 that represents (simulates) a physical simulation can be prepared as a result of model creation (training).

[0051] It is preferable that training data be prepared comprehensively for each machine learning model 171, without bias toward a large amount of training data for specific product numbers or specific product conditions. For example, if there are 10 levels of product numbers (10 different product numbers), first and second condition items belonging to the product conditions, the first condition item has three levels (three different values ​​or value ranges), and the second condition item has two levels, there are 10 × 3 × 2 = 60 combinations of product numbers and condition items, and it is preferable that the amount of training data be prepared equally for each of the 60 combinations. Furthermore, for specific product numbers that meet specific conditions (e.g., important product numbers, product numbers that exhibit complex behavior that is difficult to predict, etc.), the amount of training data may be increased compared to product numbers that do not meet such specific conditions, thereby expecting a relative improvement in prediction accuracy for the specific product numbers.

[0052] FIG. 5 shows an example of model creation (learning).

[0053] The learning unit 181 learns the machine learning model 171 using a teacher data set 172 corresponding to the machine learning model 171. Specifically, the learning unit 181 learns the machine learning model 171 for thermal stress analysis using many pairs of input data (temperature distribution data) and output data (stress analysis data) in the teacher data set 172, which is correct data corresponding to the input data. As a result, a surrogate model 122 that represents a physical simulation of the thermal stress analysis is created.

[0054] Figure 6 shows an example of model utilization (inference).

[0055] The inference unit 132 inputs temperature distribution data (data representing the temperature distribution inside the honeycomb product to be designed) as product conditions determined in a meeting between the analyst 201 and the designer 202 into the created surrogate model 122, and predicts the stress distribution that will exist inside the honeycomb product if the honeycomb product has the temperature distribution represented by the temperature distribution data.

[0056] FIG. 7 shows an example of the training data set 172.

[0057] The configuration of the teacher dataset 172 depends on the learning target. The teacher dataset 172 is made up of N teacher data 700. "N" may be any natural number, for example, the number of calculation cases in the original physical simulation. In other words, if the physical simulation is performed for N cases, the number of teacher data 700 may be N. N may be, for example, 1000 to 2000. The teacher dataset 172 illustrated in FIG. 7 is a teacher dataset used in learning to construct a surrogate model 122 that takes temperature distribution as input and stress distribution as output.

[0058] In the training data set 172, each training data 700 includes a set of input data 701 and output data 702. Take one training data 700 as an example.

[0059] The input data 701 includes a level number, a learning point number set, a position coordinate set, and an input physical quantity set. The level number corresponds to the case identification number. The learning point number set is composed of identification numbers of multiple learning points defined for a honeycomb product (typically, the interior of the honeycomb product). (In FIG. 7, "i" represents the number of learning points and may be a number independent of N.) The position coordinate set is composed of position coordinates for each learning point number in the learning point number set, and each position coordinate is the coordinate of a learning point position. The input physical quantity set is a set of input physical quantities, and includes one or more input physical quantities for each position coordinate in the position coordinate set. In the example shown in FIG. 7, the one or more input physical quantities include material stiffness (D matrix) and thermal strain (e.g., temperature and CTE (coefficient of thermal expansion)). The input data 701 may be all or a part of the input data for the physics simulation. That is, the input data 701 may be the input data for the physics simulation itself, or it may be data in which some data has been thinned out from the input data.

[0060] The output data 702 includes an output physical quantity set for the level ID, learning point number set, and position coordinate set in the input data 701. The output data 702 may include the same level ID, learning point number set, and position coordinate set as the level ID, learning point number set, and position coordinate set in the input data 701. The output physical quantity set is a set of physical quantities as output, and includes one or more types of output physical quantities for each position coordinate in the position coordinate set. In the example shown in FIG. 7, the one or more types of output physical quantities include stress. The output data 702 may be all or a part of the output data of the physics simulation. In other words, the output data 701 may be the output data of the physics simulation itself, or may be data in which some data has been thinned out from the output data.

[0061] While a physical simulation (e.g., finite element analysis) requires approximately 30 minutes per case (one level), calculations using the surrogate model 122, which is a machine learning model that has learned the behavior of the physical simulation, are expected to require approximately 10 seconds. Note that "one case" (one level) here corresponds to a set of part number, time, and product condition. For example, if calculations are performed for 1,000 product conditions and 200 times for one part number, training data 700 for 200,000 (= 1,000 × 200) cases may be prepared for that one part number, or training data 700 for a portion of those cases (e.g., 200 cases) may be prepared. In other words, there is a rule in the relationship between product conditions (e.g., temperature distribution) and product characteristics (e.g., stress distribution). As long as there is training data sufficient to learn that rule, training data for all times and / or all product conditions is not necessarily required. Similarly, for example, if calculations are performed for 1,000 product conditions for each of 200 part numbers at a single time, training data 700 for 200,000 (= 1,000 × 200) cases may be prepared for that single time, or training data 700 for a portion of those cases (e.g., 200 cases) may be prepared. In other words, a surrogate model 122 trained using training data for a portion of the part numbers can predict product characteristics for more than that portion of the part numbers. Specifically, for example, a surrogate model 122 trained using training data for 500 honeycomb product parts with different rib thicknesses, cell densities, and materials can predict (calculate) product characteristics for more than 1,000 part numbers whose design variables, such as rib thickness, cell density, and material, are within the interpolation range of the training data for the 500 part numbers. For example, in FIG. 7 , input data 701 in one training data set 700 includes data representing one temperature distribution for one part number at one time.

[0062] Each teacher data 700 in the teacher data set 172 may include a physical quantity in the teacher data 700 corresponding to a previous time depending on the type of physical quantity included in the teacher data. Specifically, for example, it is as follows. For each time, the stress (stress distribution) depends on the temperature (temperature distribution) at that time, regardless of the stress at the time immediately prior to that time. Therefore, in learning using the teacher dataset 172, the time series change in stress does not need to be a learning element. Meanwhile, the temperature (temperature distribution) at each time is affected by the temperature at the time immediately preceding that time. That is, for each position, the temperature at each time is the temperature after the change from the temperature at that position at the time immediately preceding that time. For this reason, in learning using the training dataset 172, the time series change in temperature is used as a learning element.

[0063] FIG. 8 shows an example of the accuracy verification result.

[0064] In evaluating stress transitions over a time series (multiple times), the evaluation results using the surrogate model 122 are sufficiently accurate in terms of the time and location at which maximum stress occurs compared to the evaluation results using physical simulation (even if the absolute values ​​of the stress do not match, the magnitude relationship is consistent).

[0065] Although one embodiment has been described above, this is merely an example for explaining the present invention, and the scope of the present invention is not limited to this embodiment. The present invention can be implemented in various other forms.

[0066] The above description can be summarized as follows, for example. The following summary may include supplementary explanations and explanations of modifications of the above description.

[0067] The ceramic product design support system performs the following (x) to (z). (x) A machine learning model (e.g., machine learning model 171) that replaces a physical simulation in which the product conditions of a honeycomb product, which is a ceramic product having a honeycomb structure, are input and the product characteristics of the honeycomb product are output is trained using a dataset (e.g., teacher dataset 172) that includes pairs of product conditions input into the physical simulation and product characteristics output from the physical simulation as a result of inputting the product conditions into the physical simulation. (y) By inputting product conditions of a target product, which is a honeycomb product to be designed, into a trained machine learning model (for example, the surrogate model 122), the product characteristics of the target product are predicted. (z) Outputting evaluation result information, which is information representing the predicted product characteristics of the target product.

[0068] This allows evaluation of honeycomb products with high accuracy and in a short time when designing the honeycomb product. The evaluation result information output in (z) may include information representing multiple product characteristics (e.g., a time series of product characteristics) output multiple times in (y), or may include information representing product characteristics that satisfy predetermined conditions among the multiple product characteristics (e.g., product characteristics selected by the user). The display illustrated in FIG. 8 may be one example of displaying the evaluation result information.

[0069] The "product characteristics" may include the distribution of physical quantities within the honeycomb product. As the physical quantity, at least one of the following (p) and (q) may be adopted. (p) At least one of the pressure, flow velocity, and flow rate of the fluid inside the honeycomb product. (q) At least one of the amount, the proportion, and the amount of change of a substance in a honeycomb product.

[0070] Furthermore, the relationship between the product conditions and the product characteristics may be at least one of the following (a) and (b). (a) The product conditions include the structure of the honeycomb product and the operating conditions of the honeycomb product, and the product characteristics include, as distributions, at least one of pressure distribution, temperature distribution, and flow velocity distribution inside the honeycomb product. (b) The product conditions include the structure of the honeycomb product and the temperature distribution inside the honeycomb product, and the product characteristics include the stress distribution inside the honeycomb product as a distribution.

[0071] By inputting the structure of a honeycomb product and the temperature distribution inside the honeycomb product, the stress distribution inside the honeycomb product can be obtained. By inputting data representing the obtained stress distribution into, for example, predetermined software, the probability distribution of defects inside the honeycomb product can be predicted, and the probability of failure of the honeycomb product can be predicted from the data representing the probability distribution of defects. That is, instead of or in addition to the maximum stress, the probability of defects being present can be predicted based on the volume of the honeycomb product occupied by the high stress field, and the probability of failure can be predicted from the distribution of the predicted probability. In this way, this can contribute to predicting the probability distribution of defects and the probability of failure. So-called low-dimensional prediction models can predict the maximum stress value based on the maximum temperature, etc., but cannot predict the stress distribution and therefore cannot predict the probability of failure.

[0072] The physical quantity may be one related to at least one of load, heat, atmosphere, thermal conduction, ion conduction, insulation, and filter pressure loss / capture. Specifically, examples of the meaning of the physical quantity are as follows: The distribution of physical quantities representing load contributes to predicting the probability distribution of defects and / or the probability of fracture in honeycomb products. The distribution of heat and / or atmospheric physical quantities affects the strength distribution of honeycomb products, and as a result, contributes to predicting the probability distribution of defects and / or the probability of breakage in honeycomb products. This is because deterioration progresses from areas with high temperature, high pH, ​​or high concentration inside honeycomb products, and the strength distribution changes depending on the distribution and range of that deterioration. The distribution of physical quantities related to thermal conduction, ionic conduction, or electrical insulation contributes to the improvement of honeycomb product design because, when a heat source, electrode, or voltage application point is asymmetric with respect to the honeycomb product, heat flux, ionic current density, or electric field occurs within the honeycomb product, which may limit its function. However, the distribution of physical quantities related to thermal conduction, ionic conduction, or electrical insulation can predict where within the honeycomb product such limiting factors will occur. · The physical quantities related to filter pressure drop / capture contribute to the improvement of honeycomb products in design because the distribution of flow and soot deposition is not uniform and determines the characteristics.

[0073] The physical simulations may include a first physical simulation that outputs a predetermined type of product characteristic (e.g., temperature distribution) and a second physical simulation that inputs a predetermined type of product condition including the predetermined type of product characteristic (e.g., temperature distribution).The machine learning models may include a first machine learning model (e.g., first surrogate model 122) that represents the first physical simulation and a second machine learning model (e.g., second surrogate model 122) that represents the second physical simulation.In (y) above, the second machine learning model may be input with a product condition including the predetermined type of product characteristic (e.g., temperature distribution) as an output of the first machine learning model.

[0074] For example, a machine learning model may be used that has been trained on time-series temperature distribution data of a honeycomb product obtained by heat transfer simulation (physical simulation) using a finite difference method, a finite element method, or a finite volume method, and for this simulation or machine learning model, the product conditions as input may include the internal structure of the honeycomb product and the combustion conditions of the honeycomb product (e.g., gas and soot deposition), and the product characteristics as output may include the temperature distribution.

[0075] Furthermore, for example, a machine learning model may be used that has been trained on stress distribution data of a honeycomb product obtained by a structural simulation (physical simulation) using the finite element method. For the simulation or machine learning model, the product conditions as input may include the internal structure of the honeycomb product and the temperature distribution of the honeycomb product, and the product characteristics as output may include stress distribution. An example of the accuracy verification result of this machine learning model (for example, the second machine learning model described above) is shown in FIG. 8. The temperature distribution data as a product condition may be data obtained from experimental results, data included in the input data for the physical simulation, or data included in the output data of another machine learning model (for example, the first machine learning model described above).

[0076] The machine learning model may be a model using a neural network or a model using a graph neural network (e.g., a graph convolutional network).

[0077] The evaluation result information may include at least one of the following (F) to (H). (F) Time series of maximum physical quantities inside the honeycomb product and time series of the locations where the maximum physical quantities occur. (G) Time series of minimum physical quantities inside the honeycomb product and time series of the locations where the minimum physical quantities occur. (H) Time series of specific physical quantities inside the honeycomb product and time series of the locations where the specific physical quantities occur.

[0078] The ceramic product design support system may perform at least one of the following (f) to (h). The following (f) to (h) may be performed, for example, by a simulation unit of the ceramic product design support system. According to the following, the time at which the physical quantity of interest was obtained using a machine learning model is identified, and therefore, as necessary, physical simulation may be performed only for that time or a time nearby. This is expected to result in a more accurate evaluation of the time of interest. (f) When the evaluation result information includes (F), for each of one or more times in a time range corresponding to the time series that includes the time at which the maximum value of the maximum physical quantity was obtained (for example, the time at which the maximum value was obtained and times nearby), a physical simulation is performed using product conditions as input corresponding to product characteristics that include the maximum value of the maximum physical quantity and the location where it occurred. (g) When the evaluation result information includes (G), for each of one or more times in a time range corresponding to the time series, including the time at which the minimum value of the minimum physical quantity was obtained, a physical simulation is performed using product conditions as input corresponding to product characteristics including the minimum value of the minimum physical quantity and the location where it occurred. (h) When the evaluation result information includes (H), for each of one or more times in a time range corresponding to the time series, including the time at which a specific value of a specific physical quantity was obtained, a physical simulation is performed using product conditions as input corresponding to product characteristics including the minimum value of the minimum physical quantity and the location where it occurred.

[0079] During the honeycomb product design process, the following may occur during meetings between designers and analysts: In (y), for example, an inference unit of a ceramic product design support system inputs product conditions determined by an analyst based on requests from a designer into a machine learning model. If the product characteristics represented by the evaluation result information output in (z) as a result of (y) (for example, the product characteristics represented by the evaluation result information output by the output unit of the ceramic product design support system as a result of prediction by the inference unit) do not satisfy the designer's requirements, another product condition is determined and (y) is performed. If the product characteristics represented by the evaluation result information output in (z) as a result of (y) satisfy the designer's requirements, the repetition of (y) and (z) is terminated.

[0080] In addition, the determination of whether or not the "product characteristics represented by the evaluation result information output in (z) as a result of (y)" satisfy the requirements of the designer may be performed by the computer by receiving input from the designer who has viewed the display of the product characteristics represented by the evaluation result information as to whether or not the requirements are satisfied.

[0081] The "product characteristics" may include the distribution of physical quantities on the surface of the honeycomb product (e.g., the end face and / or side face of a columnar honeycomb product) in addition to the distribution of physical quantities inside the honeycomb product. Specifically, for example, the position coordinate set in the training data 700 may include the coordinates of positions on the surface of the honeycomb product in addition to the coordinates of positions inside the honeycomb product. [Explanation of symbols]

[0082] 50...ceramics product design support system, 121...simulation model, 122...trained model, 131...simulation unit, 132...inference unit, 133...output unit, 181...learning unit

Claims

1. (x) A machine learning model that replaces a physical simulation in which product conditions of a honeycomb product, which is a ceramic product having a honeycomb structure, are input and product characteristics of the honeycomb product are output is trained using a data set that includes pairs of the product conditions input into the physical simulation and the product characteristics output from the physical simulation as a result of inputting the product conditions into the physical simulation; (y) predicting the product characteristics of a target product, which is a honeycomb product to be designed, by inputting product conditions into the trained machine learning model; (z) outputting evaluation result information that is information representing predicted product characteristics of the target product; Ceramic product design support method.

2. The product characteristics include a distribution of physical quantities within the honeycomb product. The ceramic product design support method according to claim 1.

3. The physical quantity is at least one of the following (p) and (q): (p) at least one of the pressure, flow velocity, and flow rate of a fluid within the honeycomb product; (q) at least one of the amount, proportion, and variation of a substance present in the honeycomb product; The ceramic product design support method according to claim 2.

4. The relationship between the product conditions and the product characteristics is at least one of the following (a) and (b): (a) the product conditions include a structure of a honeycomb product and an operating condition of the honeycomb product, and the product characteristics include, as the distribution, at least one of a pressure distribution, a temperature distribution, and a flow velocity distribution inside the honeycomb product; (b) the product conditions include a structure of the honeycomb product and a temperature distribution inside the honeycomb product, and the product characteristics include a stress distribution inside the honeycomb product as the distribution; The ceramic product design support method according to claim 2.

5. The machine learning model is a model using a neural network or a model using a graph neural network. The ceramic product design support method according to claim 1.

6. The physical simulations include a first physical simulation that outputs a predetermined type of product characteristic, and a second physical simulation that inputs a predetermined type of product condition including the predetermined type of product characteristic, the machine learning models include a first machine learning model representing the first physical simulation and a second machine learning model representing the second physical simulation; In the step (y), product conditions including predetermined types of product characteristics as an output of the first machine learning model are input to the second machine learning model. The ceramic product design support method according to claim 1.

7. The evaluation result information includes at least one of the following (F) to (H): (F) a time series of the maximum physical quantity inside the honeycomb product and a time series of the location where the maximum physical quantity occurs; (G) a time series of the minimum physical quantity inside the honeycomb product and a time series of the location where the minimum physical quantity occurs; (H) a time series of a specific physical quantity inside the honeycomb product and a time series of the location of the specific physical quantity; The ceramic product design support method according to claim 1.

8. Perform at least one of the following (f) to (h): (f) when the evaluation result information includes (F), for each of one or more times in a time range corresponding to the time series, including the time at which a maximum value of a maximum physical quantity is obtained, the physical simulation is performed using product conditions as input corresponding to product characteristics including the maximum value of the maximum physical quantity and the location where the maximum value occurs. (g) when the evaluation result information includes (G), for each of one or more times in a time range corresponding to the time series, including the time at which the minimum value of the minimum physical quantity was obtained, the physical simulation is performed using product conditions as input corresponding to product characteristics including the minimum value of the minimum physical quantity and the location where the minimum value occurred. (h) when the evaluation result information includes (H), for each of one or more times in a time range corresponding to the time series, including a time at which a specific value of a specific physical quantity was obtained, the physical simulation is performed using product conditions as input corresponding to product characteristics including the minimum value of the minimum physical quantity and the location where the minimum value occurred. The ceramic product design support method according to claim 7.

9. During the honeycomb product design process, a meeting was held between the designer and the analyst. (y) inputting product conditions determined by the analyst based on a request from the designer into the machine learning model; If the product characteristics represented by the evaluation result information output in (z) as a result of (y) do not satisfy the requirements of the designer, another product condition is determined and (y) is performed; If the product characteristics represented by the evaluation result information output in (z) as a result of (y) satisfy the requirements of the designer, the repetition of (y) and (z) is terminated. The ceramic product design support method according to claim 1.

10. a learning unit that learns a machine learning model that replaces a physical simulation in which product conditions of a honeycomb product, which is a ceramic product having a honeycomb structure, are input and product characteristics of the honeycomb product are output, using a data set that includes pairs of the product conditions input to the physical simulation and the product characteristics output from the physical simulation as a result of inputting the product conditions to the physical simulation; an inference unit that predicts product characteristics of a target product, which is a honeycomb product to be designed, by inputting product conditions of the target product into the trained machine learning model; an output unit that outputs evaluation result information that is information representing predicted product characteristics of the target product; A ceramic product design support system equipped with:

Images