Inference program, inference method, and information processing device

The use of trained models to generate coarse mesh data from fine mesh data in simulation analysis addresses the computational burden and time issues, enabling high-speed and accurate inference of simulation results.

JP7590653B2Active Publication Date: 2024-11-27FUJITSU LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
JP2021063522
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2021-04-02
Publication Date
2024-11-27
Estimated Expiration
2041-04-02

AI Technical Summary

Technical Problem

Existing simulation methods using fine mesh data for analysis result in high computational load and long execution times, while the creation of corresponding coarse mesh data is often overlooked, leading to increased burden and slow inference.

Method used

An inference program utilizing trained models to remove edges from mesh data, generating coarse mesh data automatically and inferring simulation results from fine mesh data, reducing computational load and execution time.

Benefits of technology

Enables high-speed inference of simulation results with improved accuracy by generating coarse mesh data automatically, reducing computational requirements and preserving essential features of fine mesh data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007590653000003
    Figure 0007590653000003
  • Figure 0007590653000004
    Figure 0007590653000004
  • Figure 0007590653000005
    Figure 0007590653000005
Patent Text Reader

Abstract

To enable fast inference of simulation results.SOLUTION: A storage unit stores a first learnt model for determining features of mesh data by removing some edges from mesh data including a plurality of nodes and edges between the nodes. A processing unit uses the first learnt model to generate, from second mesh data, first mesh data having fewer edges than the second mesh data. The processing unit executes a simulation using the first mesh data to generate first simulation result data indicating the physical quantity of an object indicated by the first mesh data. The processing unit uses a second learned model to infer, from the first mesh data and the first simulation result data, second simulation result data corresponding to a case where the simulation is performed using the second mesh data.SELECTED DRAWING: Figure 8
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to an inference program, an inference method, and an information processing device. [Background technology]

[0002] Simulations are performed to analyze the physical properties of an object based on its shape. Simulations include structural analysis to calculate stress and displacement, fluid analysis to calculate flow velocity and fluid density, and electromagnetic field analysis to calculate electric current and magnetic force. Simulations may use mesh data including discrete nodes and edges connecting the nodes as shape data that indicates the shape of the object. Simulations may also use numerical analysis methods such as the finite element method (FEM) to obtain approximate solutions to governing equations that represent physical phenomena. For example, numerical analysis methods assign variables to nodes, generate coefficient matrices that indicate the relationships between the nodes, and obtain approximate solutions to simultaneous equations using an iterative method.

[0003] A resolution control system has been proposed that integrates vertex pairs included in a high-resolution mesh model to generate a low-resolution mesh model suitable for finite element analysis.A mesh correction device has also been proposed that uses a neural network to group triangular meshes included in a mesh model, and reflects corrections to a certain triangular mesh in other triangular meshes in the same group.A mesh generation device has also been proposed that generates a response surface from a representative shape model and parameters indicating mesh density. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2005-242647 A [Patent Document 2] JP 2006-318232 A [Patent Document 3] JP 2011-243016 A Summary of the Invention [Problem to be solved by the invention]

[0005] By using fine mesh data with a large element area, the simulation can generate highly accurate analysis results, but the amount of calculation is large and the execution time is long. Therefore, a method can be considered in which a simulation is performed using coarse mesh data with a small element area, and the fine simulation results when the fine mesh data is used are inferred from the coarse simulation results. A trained model generated by machine learning, such as a neural network, may be used for the inference.

[0006] However, when performing a simulation, although fine mesh data has been created by a user, coarse mesh data corresponding to the fine mesh data may not have been created. For example, when a user designs an industrial part and checks its physical properties by simulation, design data corresponding to a fine mesh model may have been created, but a coarse mesh model may not have been created.

[0007] Therefore, the burden of creating a coarse mesh model is generated, and it may be difficult to infer the simulation results at high speed. Therefore, in one aspect, the present invention aims to enable inference of the simulation results to be performed at high speed. [Means for solving the problem]

[0008] In one embodiment, an inference program is provided that causes a computer to execute the following processes: Generate, from second mesh data, first mesh data having fewer edges than the second mesh data, using a first trained model that removes some edges from mesh data including a plurality of nodes and edges between the nodes to determine a feature of the mesh data; Run a simulation using the first mesh data to generate first simulation result data indicating a physical quantity of an object indicated by the first mesh data; Infer, from the first mesh data and the first simulation result data, second simulation result data corresponding to a case where a simulation is run using the second mesh data, using the second trained model.

[0009] In one embodiment, a computer-implemented inference method is provided.In one embodiment, an information processing device is provided having a storage unit and a processing unit. Effect of the Invention

[0010] In one aspect, inference of simulation results can be performed quickly. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating an information processing device according to a first embodiment. [Diagram 2] FIG. 11 illustrates an example of hardware of an information processing apparatus according to a second embodiment. [Diagram 3] FIG. 13 is a diagram showing an example of stress prediction using a neural network. [Figure 4] FIG. 1 is a diagram showing examples of a fine mesh model and a coarse mesh model. [Diagram 5] FIG. 1 is a diagram illustrating an example of the structure of a neural network. [Figure 6] FIG. 13 is a diagram illustrating example parameters of corners and fillets. [Figure 7] 13 is a graph showing an example of a speed multiplication factor for stress prediction; [Figure 8] FIG. 13 is a diagram showing a second example of stress prediction using a neural network. [Figure 9] FIG. 1 is a diagram illustrating an example of the structure of a graph convolutional neural network. [Figure 10] FIG. 1 is a diagram illustrating an example of input and output of a graph convolutional neural network. [Figure 11] FIG. 13 is a diagram illustrating an example of stepwise pooling for a mesh model. [Figure 12] FIG. 11 is a diagram illustrating an example of parameters of an edge feature amount. [Figure 13] FIG. 13 is a diagram illustrating an example of edge removal by pooling. [Figure 14] FIG. 2 is a block diagram showing an example of functions of the information processing device; [Figure 15] 1A and 1B are diagrams illustrating examples of the structures of mesh data and training data. [Figure 16] 1 is a flowchart showing an example of a machine learning procedure. [Figure 17] 11 is a flowchart illustrating an example of a procedure for a stress simulation. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] The present embodiment will be described below with reference to the drawings. [First embodiment] A first embodiment will be described.

[0013] FIG. 1 is a diagram illustrating an information processing apparatus according to a first embodiment. The information processing device 10 executes a simulation to analyze the physical characteristics of an object from the shape of the object. The simulation may be a structural analysis to calculate stress or displacement, a fluid analysis to calculate flow velocity or fluid density, or an electromagnetic field analysis to calculate current or magnetic force. The information processing device 10 may be a client device or a server device. The information processing device 10 may be called a computer, an analysis device, or a simulation device.

[0014] The information processing device 10 has a storage unit 11 and a processing unit 12. The storage unit 11 may be a volatile semiconductor memory such as a random access memory (RAM), or may be a non-volatile storage such as a hard disk drive (HDD) or a flash memory. The processing unit 12 is a processor such as a central processing unit (CPU), a graphics processing unit (GPU), or a digital signal processor (DSP). However, the processing unit 12 may include an electronic circuit for a specific purpose such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). The processor executes a program stored in a memory such as a RAM. A collection of processors may be called a multiprocessor or simply a "processor."

[0015] The memory unit 11 stores trained models 13 and 14 and mesh data 16. The trained model 13 is a judgment model that judges the characteristics of mesh data. Mesh data is shape data that indicates the shape of an object, and includes a plurality of nodes and a plurality of edges that connect the nodes. Mesh data may be called graph data. The trained model 13 judges the characteristics of the mesh data by removing some edges from the mesh data to simplify the data. For example, the trained model 13 judges the type (class) of an object indicated by the mesh data. The trained model 13 is generated through machine learning.

[0016] The trained model 13 may be a neural network or a graph convolutional neural network that performs a convolution operation on mesh data. For example, the trained model 13 performs one or more of a convolution operation that calculates the feature amount of each edge based on the positional relationship with adjacent edges, and a pooling operation that removes some edges based on the feature amount of each edge. The trained model 13 may also include multiple pooling layers that gradually reduce the number of edges. For example, the trained model 13 includes a specific number of pairs of a convolution layer that performs a convolution operation and a pooling layer that performs a pooling operation.

[0017] The trained model 14 is an inference model that infers the physical quantity of an object represented by mesh data. In the case of structural analysis, the physical quantity is, for example, the stress or displacement amount at a critical portion where large stress is likely to occur, such as a corner or a hole of the object. The trained model 14 is generated through machine learning. The trained model 14 may be a neural network.

[0018] The trained model 14, for example, infers physical quantities calculated on fine mesh data with many edges from parameters related to the shape represented by coarse mesh data with few edges and parameters related to physical quantities calculated on the coarse mesh data. It can also be said that the trained model 14 is a conversion model that converts coarse simulation result data into highly accurate simulation result data. The trained model 14, for example, infers stress occurring in a fillet from parameters related to the shape of an acute angle represented by few edges and parameters related to stress occurring in the acute angle.

[0019] The mesh data 16 is fine mesh data having more edges than the mesh data 15 described below. The mesh data 16 includes, for example, fillets with rounded corners so as to have no vertices. The fillet shape uses more edges than a sharp corner shape. The mesh data 16 may be generated from design data of an industrial part, or may be generated from CAD (Computer Aided Design) data by mesh generation software.

[0020] The processing unit 12 uses the trained models 13 and 14 to infer simulation result data 18 corresponding to the mesh data 16 without directly executing a simulation on the mesh data 16. The simulation result data 18 includes physical quantities of the object represented by the mesh data 16, for example, stress in a fillet. The simulation may execute a numerical analysis method such as a finite element method. For example, the simulation assigns variables to nodes, generates a coefficient matrix indicating the relationship between the nodes, and calculates an approximate solution to a simultaneous equation by an iterative method that iteratively performs matrix calculations.

[0021] First, the processing unit 12 uses the trained model 13 to generate mesh data 15 having fewer edges than the mesh data 16 from the mesh data 16. The reduction in edges may convert the shape of a fillet into a shape with an acute angle that is sharper than the fillet. At this time, the processing unit 12 may extract intermediate data generated inside the trained model 13 by determining the characteristics of the mesh data 16. The processing unit 12 may also extract the output of a pooling layer included in the trained model 13. The processing unit 12 may extract the output of an intermediate pooling layer among a plurality of pooling layers. For example, the processing unit 12 extracts the output of the second or third pooling layer among the four pooling layers.

[0022] When the mesh data 15 is generated, the processing unit 12 executes a simulation such as a finite element method using the mesh data 15 to generate simulation result data 17 corresponding to the mesh data 15. The simulation result data 17 includes physical quantities of the object represented by the mesh data 15, for example, stress at an acute angle. Since the simulation for the mesh data 15 has fewer edges than the simulation for the mesh data 16, the amount of calculation is smaller and the execution time is shorter.

[0023] Then, the processing unit 12 uses the trained model 14 to infer simulation result data 18 from the mesh data 15 and the simulation result data 17. For example, the processing unit 12 generates input data including parameters corresponding to the mesh data 15 and parameters corresponding to the simulation result data 17, inputs the input data to the trained model 14, and acquires an inferred value output by the trained model 14.

[0024] As described above, the information processing device 10 of the first embodiment converts mesh data 16 having many edges into mesh data 15 having few edges by using the trained model 13 that removes some edges and determines the characteristics of the mesh data. The information processing device 10 performs a simulation using the mesh data 15. Then, the information processing device 10 infers simulation result data 18 corresponding to the mesh data 16 from the mesh data 15 and simulation result data 17 corresponding to the mesh data 15.

[0025] This allows the information processing device 10 to generate highly accurate simulation result data 18 corresponding to the fine mesh data 16. Furthermore, a simulation using the coarse mesh data 15 requires less calculation and is faster than a simulation using the fine mesh data 16. Therefore, the amount of calculation required to obtain the highly accurate simulation result data 18 is reduced, and the execution time is shortened.

[0026] Furthermore, the information processing device 10 automatically generates coarse mesh data 15 corresponding to the fine mesh data 16 by using the trained model 13. This reduces the burden of creating the mesh data 15 separately from the mesh data 16, and the information processing device 10 can quickly execute inference of the simulation result data 18. For example, when a user is designing the shape of an industrial part, the information processing device 10 can execute inference of the simulation result data 18 if design data exists.

[0027] Moreover, the trained model 13 is trained to remove edges that are less related to features in order to determine the features of the mesh data. Therefore, the mesh data 15 generated using the trained model 13 is a good coarse mesh model that preserves the essential features of the mesh data 16. As a result, the accuracy of the simulation result data 17 is improved, and the inference accuracy of the simulation result data 18 is improved.

[0028] [Second embodiment] Next, a second embodiment will be described. The information processing device 100 of the second embodiment executes a structural analysis simulation using a finite element method, and calculates stress according to the shape of an object. Moreover, the information processing device 100 predicts a highly accurate simulation result using machine learning. The information processing device 100 is used, for example, in the design of industrial parts. The information processing device 100 may be a client device or a server device. The information processing device 100 may be called a computer, an analysis device, a simulation device, a design device, or a machine learning device.

[0029] FIG. 2 illustrates an example of hardware of an information processing apparatus according to the second embodiment. The information processing device 100 has a CPU 101, a RAM 102, a HDD 103, a GPU 104, an input interface 105, a medium reader 106, and a communication interface 107, all connected via a bus. The CPU 101 corresponds to the processing unit 12 in the first embodiment. The RAM 102 or the HDD 103 corresponds to the storage unit 11 in the first embodiment.

[0030] The CPU 101 is a processor that executes instructions of a program. The CPU 101 loads at least a part of the program and data stored in the HDD 103 into the RAM 102 and executes the program. The information processing device 100 may have multiple processors. A collection of processors may be called a multiprocessor or simply a "processor."

[0031] The RAM 102 is a volatile semiconductor memory that temporarily stores programs executed by the CPU 101 and data used in calculations by the CPU 101. The information processing device 100 may have a type of volatile memory other than a RAM.

[0032] The HDD 103 is a non-volatile storage that stores software programs such as an OS (Operating System), middleware, and application software, and data. The information processing device 100 may have other types of non-volatile storage, such as a flash memory or an SSD (Solid State Drive).

[0033] The GPU 104 performs image processing in cooperation with the CPU 101. The GPU 104 also outputs an image to a display device 111 connected to the information processing device 100. The display device 111 is, for example, a CRT (Cathode Ray Tube) display, a liquid crystal display, an organic EL (Electro Luminescence) display, or a projector. Note that other types of output devices such as a printer may be connected to the information processing device 100.

[0034] The input interface 105 receives an input signal from an input device 112 connected to the information processing device 100. The input device 112 is, for example, a mouse, a touch panel, or a keyboard. A plurality of input devices may be connected to the information processing device 100.

[0035] The medium reader 106 is a reading device that reads the program and data recorded on the recording medium 113. The recording medium 113 is, for example, a magnetic disk, an optical disk, or a semiconductor memory. The magnetic disk includes a flexible disk (FD) and a HDD. The optical disk includes a compact disc (CD) and a digital versatile disc (DVD). The medium reader 106 copies the program and data read from the recording medium 113 to another recording medium such as the RAM 102 or the HDD 103. The read program may be executed by the CPU 101.

[0036] The recording medium 113 may be a portable recording medium. The recording medium 113 may be used for distributing programs and data. The recording medium 113 and the HDD 103 may be called computer-readable recording media.

[0037] The communication interface 107 is connected to a network 114 and communicates with other information processing devices via the network 114. The communication interface 107 may be a wired communication interface connected to a wired communication device such as a switch or a router, or a wireless communication interface connected to a wireless communication device such as a base station or an access point.

[0038] Next, the flow of the structural analysis simulation will be described. FIG. 3 is a diagram showing an example of stress prediction using a neural network. In the second embodiment, a corner of an industrial part is assumed as a part for which stress is to be calculated with high accuracy. A sharp angle with an angle between 0 degrees and 180 degrees and a vertex is sometimes called a "corner." When an external force is applied to an industrial part that includes a corner, stress concentration may occur in which the stress in a specific part is significantly greater than in other parts. Therefore, when designing an industrial part, the shape of the corner may be made a fillet rather than a corner in order to alleviate stress concentration. A fillet is a smooth corner that does not have a vertex and has a curve with a constant radius of curvature. The stress near the fillet is expected to be smaller than that of a corner.

[0039] As one method for calculating the stress of the fillet, a fine mesh model 31 that represents the shape of the fillet with high accuracy is created, and a finite element method simulation is performed using the fine mesh model 31. The mesh model is a shape model that represents the shape of an object with a mesh, and includes a plurality of discrete nodes and a plurality of edges that connect the nodes. In the mesh model, the space in which the object exists is divided into elements that are small regions surrounded by the nodes and edges. In the second embodiment, the faces of the elements are triangular. Usually, CAD data exists for the industrial parts being designed. Therefore, the information processing device 100 generates the fine mesh model 31 from the CAD data using, for example, mesh generation software.

[0040] However, a finite element method simulation using a fine mesh model 31 with many nodes and edges may require a large amount of calculation and a long execution time. Therefore, the information processing device 100 executes a finite element method simulation using a coarse mesh model 32 instead of the finite element method simulation using the fine mesh model 31.

[0041] First, a coarse mesh model 32 corresponding to the fine mesh model 31 is created. The coarse mesh model 32 has fewer nodes and edges than the fine mesh model 31. Therefore, the accuracy of expressing the shape of the coarse mesh model 32 is lower than that of the fine mesh model 31. As a result, it is difficult for the coarse mesh model 32 to sufficiently reproduce the curve of a fillet, and the shape of a corner may be expressed as a corner rather than a fillet.

[0042] The information processing device 100 executes a finite element method simulation using the coarse mesh model 32 to generate a coarse simulation result 33. For example, the information processing device 100 assigns variables to the nodes of the coarse mesh model 32, and generates a coefficient matrix indicating the relationship between the nodes based on governing equations indicating the physical laws of structural mechanics. The information processing device 100 uses the coefficient matrix to obtain an approximate solution to the simultaneous equations by an iterative method. The coarse simulation result 33 includes stress near the corner when the corner shape is a corner.

[0043] The information processing device 100 generates input data for a neural network 140 that has been generated in advance by machine learning. The input data is a tensor including a feature amount related to stress indicated by the coarse simulation result 33, a feature amount related to a corner shape indicated by the coarse mesh model 32, and a feature amount related to a fillet shape indicated by the fine mesh model 31. The neural network 140 predicts stress near the fillet. As a result, a fine simulation result 34 including a predicted value of the stress of the fillet is generated.

[0044] The stress indicated by the fine simulation result 34 corresponds to the stress calculated when a finite element method simulation is performed using the fine mesh model 31. Therefore, the accuracy of the fine simulation result 34 is higher than that of the coarse simulation result 33. On the other hand, the amount of calculation of the finite element method simulation using the coarse mesh model 32 is smaller than that of the fine mesh model 31. In addition, the amount of calculation for prediction by the neural network 140 is sufficiently small. Therefore, the execution time of the structural analysis simulation is shortened.

[0045] Here, the highly accurate stress may be calculated by focusing on critical parts such as corners where stress concentration is likely to occur. For stresses other than critical parts, stresses calculated from the coarse mesh model 32 may be used. Therefore, the information processing device 100 calculates the stress distribution of the entire object using the coarse mesh model 32, and converts the stresses of the corners of the entire object into fillet stresses using the neural network 140. The neural network 140 calculates a predicted value of stress for each fillet.

[0046] FIG. 4 is a diagram showing an example of a fine mesh model and a coarse mesh model. The fine mesh model 35 corresponds to the above-mentioned fine mesh model 31. The coarse mesh model 36 corresponds to the above-mentioned coarse mesh model 32. The fine mesh model 35 and the coarse mesh model 36 represent the shape of the same industrial part. In the fine mesh model 35 and the coarse mesh model 36, all of the faces of the elements are triangles.

[0047] This industrial part has two holes and one notch. This industrial part also has ten corners on the outside of the part whose angles are greater than 0 degrees and less than 180 degrees. Of these ten corners, one corner is 45 degrees, one corner is 60 degrees, six corners are 90 degrees, one corner is 120 degrees, and one corner is 150 degrees. In the fine mesh model 35, the shape of the corner is a fillet. In the coarse mesh model 36, the shape of the corner is a corner. Therefore, the fine mesh model 35 has a significantly larger number of nodes and edges near the corners than the coarse mesh model 36.

[0048] FIG. 5 is a diagram illustrating an example of the structure of a neural network. Neural network 140 includes an input layer 141, a hidden layer 142, and an output layer 143. Hidden layer 142 is located between input layer 141 and output layer 143. Input layer 141 includes four nodes. Hidden layer 142 includes ten nodes. Output layer 143 includes one node. Each node of input layer 141 is connected to ten nodes of hidden layer 142 by edges. Each node of hidden layer 142 is connected to a node of output layer 143.

[0049] The four nodes in the input layer 141 correspond to the corner angle 41, the mesh size 42, the fillet radius 43, and the stress ratio 44. One node in the output layer corresponds to the fillet stress 45. The corner angle 41, the mesh size 42, the fillet radius 43, and the stress ratio 44 are input data corresponding to explanatory variables. The fillet stress 45 is output data corresponding to the objective variable. In the machine learning that generates the neural network 140, the output fillet stress 45 is compared with teacher data indicating the correct fillet stress.

[0050] The corner angle 41 is the angle of one corner included in the coarse mesh model 32. The mesh size 42 is the size of the mesh in the coarse mesh model 32 near the corner. The fillet radius 43 is the radius of curvature of the fillet included in the fine mesh model 31 corresponding to the corner. The stress ratio 44 is the ratio of the magnitudes of stresses acting in different directions on either side of the apex of the corner. For example, the stress ratio 44 is the ratio of the magnitude of the horizontal stress near the corner to the magnitude of the vertical stress on the opposite side of the apex.

[0051] The corner angle 41 and the mesh size 42 correspond to the features of the coarse mesh model 32. The fillet radius 43 corresponds to the features of the fine mesh model 31. The stress ratio 44 corresponds to the features of the coarse simulation result 33. The fillet stress 45 is a predicted value of the stress in the fillet corresponding to the corner.

[0052] Weights are assigned to edges included in the neural network 140. In generating output data, the product of the output value of a node and the weight of the edge is transmitted along the edge to the subsequent node. The subsequent node inputs the sum of the received products into a predetermined activation function to calculate its own output value. In this way, output data is generated from input data by forward processing from the input layer 141 to the output layer 143.

[0053] In machine learning, edge weights are determined by backpropagation. Backpropagation evaluates the error between the output of the neural network 140 and the teacher data, and calculates an error gradient that indicates the amount of change in error when the weight of each edge is changed by a small amount. Backpropagation updates the weight of each edge based on the calculated error gradient and a predetermined learning rate. The edge weights are updated in sequence by backward processing in the reverse direction from the output layer 143 to the input layer 141. That is, error information is propagated in the opposite direction to the forward processing.

[0054] The information processing device 100 generates training data including input data and teacher data, and determines weights of the neural network 140 using the generated training data. To generate the training data, the information processing device 100 lists candidate values ​​for each of the corner angle 41, mesh size 42, fillet radius 43, and stress ratio 44. The information processing device 100 generates a plurality of simulation conditions with different combination patterns of the corner angle 41, mesh size 42, fillet radius 43, and stress ratio 44. For each simulation condition, the information processing device 100 generates a sample of a fine mesh model that satisfies the condition, executes a finite element method simulation, and calculates a correct answer for the fillet stress 45.

[0055] FIG. 6 is a diagram showing example parameters of the corners and fillets. Mesh model 37 is a sample showing one corner or fillet. The length of the left side is 50 mm, the length of the bottom side is 50 mm, the length of the top side is 20 mm, and the length of the right side is 20 mm. The length between the top side and the corner vertex is 30 mm, and the length between the right side and the corner vertex is 30 mm.

[0056] Candidate values ​​for the corner angle 41 are 30 degrees, 60 degrees, 90 degrees, 120 degrees, and 150 degrees. Candidate values ​​for the mesh size 42 are 1 millimeter, 2 millimeters, 3 millimeters, 4 millimeters, and 5 millimeters. The mesh size 42 corresponds, for example, to the length of one side of an element. Candidate values ​​for the fillet radius 43 are 0.1 millimeter, 0.2 millimeter, 0.3 millimeter, 0.4 millimeter, and 0.5 millimeter. The larger the fillet radius 43, the more gentle the corner curve.

[0057] Candidate values ​​for the stress ratio 44 are 1:0, 1:0.25, 1:0.5, 1:0.75, and 1:1. The stress ratio 44 is a loading condition that indicates the pattern of forces applied to a corner. In general, the stress ratio 44 is calculated by dividing the nodal Mises stress S A ,S B It is the ratio of the smaller value to the larger value of F1 / F2. In the second embodiment, the stress ratio 44 is defined as F1 / F2. F1 is the force generated from the upper side to the lower side on the right side of the corner. F2 is the force generated from the right side to the left side on the upper side of the corner, and is perpendicular to F1. In addition, as material parameters indicating the elasticity of the object, Young's modulus is set to 70000 and Poisson's ratio is set to 0.3.

[0058] In this way, when generating training data, the information processing device 100 sets five candidate values ​​for each of the corner angle 41, mesh size 42, fillet radius 43, and stress ratio 44. The information processing device 100 generates 625 simulation conditions by exhaustively generating combination patterns of these candidate values. The information processing device 100 executes a finite element method simulation of a fine mesh model according to each of the 625 simulation conditions, and calculates the correct fillet stress. However, instead of using the finite element method simulation, a prototype part may be manufactured to measure the fillet stress.

[0059] The information processing device 100 generates the neural network 140 by machine learning based on the generated training data. By predicting the microscopic simulation result 34 using the neural network 140, the speed of the structural analysis simulation is increased.

[0060] FIG. 7 is a graph showing an example of a speed multiplier for stress prediction. The horizontal axis of graph 91 represents the mesh size of the coarse mesh model. The vertical axis of graph 91 represents the speed magnification. The numerator of the speed magnification is the execution time of the finite element method simulation using the fine mesh model. The denominator of the speed magnification is the sum of the execution time of the finite element method simulation using the coarse mesh model and the execution time of the fillet stress prediction by the neural network 140. The larger the speed magnification, the greater the effect of reducing the execution time.

[0061] As shown in graph 91, the speed multiplication factor is about 1.7 times when the mesh size is 1 mm. The speed multiplication factor is about 1.8 times when the mesh size is 2 mm. The speed multiplication factor is about 1.9 times when the mesh size is 3 mm. The speed multiplication factor is about 2.2 times when the mesh size is 4 mm. The speed multiplication factor is about 2.88 times when the mesh size is 5 mm. In this way, the fillet stress prediction using the neural network 140 is faster than the numerical analysis method using only the finite element method simulation. Also, the larger the mesh size, the greater the effect of the speed increase.

[0062] Next, the creation of the coarse mesh model 32 will be described. FIG. 8 is a diagram showing a second example of stress prediction using a neural network. When a fine mesh model 31 exists first, a coarse mesh model 32 is created based on the fine mesh model 31. However, it is a heavy burden for a user to manually create the coarse mesh model 32. In particular, when an industrial part includes many fillets, the correction work of correcting each fillet into a corner is a heavy burden. In addition, the mesh pattern of the coarse mesh model 32 corresponding to the fine mesh model 31 is not necessarily obvious from the fine mesh model 31. Therefore, there is a risk that the quality of the coarse mesh model 32 will be low.

[0063] Therefore, the information processing device 100 automatically generates the coarse mesh model 32 from the fine mesh model 31 by using a graph convolutional neural network 150. The graph convolutional neural network 150 is a convolutional neural network (CNN) generated by machine learning. The convolutional neural network is a multi-layer neural network that performs convolution operations, which are a type of product-sum operations.

[0064] However, unlike a general convolutional neural network, the graph convolutional neural network 150 handles mesh data instead of image data. A general convolutional neural network calculates the feature amounts of pixels arranged in a lattice according to the adjacent relationships between the pixels. On the other hand, the graph convolutional neural network 150 calculates the feature amounts of edges included in a mesh model according to the connection relationships between the edges. The graph convolutional neural network 150 may be called a mesh convolutional neural network.

[0065] The information processing device 100 inputs edge information included in the fine mesh model 31 to the graph convolutional neural network 150. The graph convolutional neural network 150 calculates feature amounts of each edge and removes some edges based on the edge feature amounts. The graph convolutional neural network 150 determines the class (type) of the object indicated by the fine mesh model 31 from the feature amounts of the remaining edges. The information processing device 100 generates the coarse mesh model 32 by extracting intermediate data generated during the class determination from the graph convolutional neural network 150.

[0066] The coarse mesh model 32 generated using the graph convolutional neural network 150 includes important edges that represent essential features of the fine mesh model 31. For example, the information processing device 100 generates a coarse mesh model 32 including 8626 elements from a fine mesh model 31 including 25724 elements.

[0067] FIG. 9 is a diagram illustrating an example of the structure of a graph convolutional neural network. The graph convolutional neural network 150 includes four convolutional layers 151-1 to 151-4, four pooling layers 152-1 to 152-4, and two fully connected layers 153-1 and 153-2. The pooling layer 152-1 is located next to the convolutional layer 151-1. The pooling layer 151-2 is located next to the convolutional layer 151-1. The pooling layer 152-2 is located next to the convolutional layer 151-2. The convolutional layer 151-3 is located next to the pooling layer 152-2. The pooling layer 152-3 is located next to the convolutional layer 151-3. The convolutional layer 151-4 is located next to the pooling layer 152-3. The pooling layer 152-4 is located next to the convolutional layer 151-4. The fully connected layer 153-1 is next to the pooling layer 152-4. The fully connected layer 153-1 is next to the fully connected layer 153-2.

[0068] The graph convolutional neural network 150 processes the connected 750 edges as one batch. The graph convolutional neural network 150 receives initial features of the 750 edges. Each edge feature is a five-dimensional vector. The graph convolutional neural network 150 also outputs a class vector indicating the class of the object indicated by the input data. The class vector is a 30-dimensional vector. One dimension of the class vector corresponds to one class. The class vector lists the probability that the object indicated by the input data belongs to each of the 30 classes. The class vector indicates that the input data is classified into the class with the maximum probability.

[0069] The convolutional layers 151-1 to 151-4 update the feature amount of each edge by a convolution operation. The feature amount of a certain edge is updated so as to reflect the feature amount of an adjacent edge adjacent to that edge. The number of edges before and after the convolutional layers 151-1 to 151-4 does not change.

[0070] The pooling layers 152-1 to 152-4 remove some edges by pooling calculation. The pooling layers 152-1 to 152-4 determine edges with low importance based on the edge feature amount calculated by the previous convolution layer, and remove the edges with low importance. The number of edges to be removed is a hyperparameter given before machine learning. When an edge is removed, the nodes at both ends of the edge are integrated. In addition, the feature amount of the adjacent edges adjacent to the edge is updated.

[0071] The convolutional layer 151-1 outputs 750 edge feature amounts. The pooling layer 152-1 removes 150 edges from the 750 edges, and outputs 600 edge feature amounts. The convolutional layer 151-2 outputs 600 edge feature amounts. The pooling layer 152-2 removes 150 edges from the 600 edges, and outputs 450 edge feature amounts. The convolutional layer 151-3 outputs 450 edge feature amounts. The pooling layer 152-3 removes 150 edges from the 450 edges, and outputs 300 edge feature amounts. The convolutional layer 151-4 outputs 300 edge feature amounts. The pooling layer 152-4 removes 120 edges from the 300 edges, and outputs 180 edge feature amounts.

[0072] The fully connected layer 153-1 generates, from 180 edge features, a vector whose number of dimensions is greater than 30. The fully connected layer 153-2 generates a 30-dimensional class vector from the vector output by the fully connected layer 153-1, and outputs the class vector.

[0073] However, the information processing device 100 does not use the class vector itself output by the graph convolutional neural network 150 in the structural analysis simulation. The information processing device 100 inputs the fine mesh model 31 to the graph convolutional neural network 150 and extracts the output of the pooling layer 152-2. The output of the pooling layer 152-2 is information indicating 450 edges that are highly important for class determination among the 750 edges. The information processing device 100 generates a set of edges indicated by the output of the pooling layer 152-2 as the coarse mesh model 32.

[0074] In the second embodiment, the number of pooling layers is four, but the number of pooling layers can be changed. In the second embodiment, the information processing device 100 extracts the output of the second pooling layer, but the output of another pooling layer such as the third pooling layer may be extracted. The pooling layer to be used may be determined during machine learning of the graph convolutional neural network 150, or may be selected when generating the crude simulation result 33. The pooling layer to be used may be specified by the user.

[0075] FIG. 10 is a diagram illustrating an example of input and output of a graph convolutional neural network. The information processing device 100 generates training data for the graph convolution neural network 150 separately from the training data for the neural network 140, and performs machine learning. The training data is generated from samples of a mesh model to which a teacher label is attached. Each sample represents the shape of an object belonging to one of 30 classes. The teacher label indicates the correct class to which the object belongs.

[0076] For example, mesh models 51, 52, and 53 are samples of mesh models. Mesh model 51 represents the shape of a dog and is assigned a teacher label indicating that the class is dog. Mesh model 52 represents the shape of a cat and is assigned a teacher label indicating that the class is cat. Mesh model 53 represents the shape of a laptop computer and is assigned a teacher label indicating that the class is laptop.

[0077] The information processing device 100 extracts edges from the mesh model 51 to generate input data, and generates teacher data 54 in which only the dimension corresponding to a dog is 1 and the other dimensions are 0. Similarly, the information processing device 100 extracts edges from the mesh model 52 to generate input data, and generates teacher data 55 in which only the dimension corresponding to a cat is 1 and the other dimensions are 0. Furthermore, the information processing device 100 extracts edges from the mesh model 53 to generate input data, and generates teacher data 56 in which only the dimension corresponding to a laptop is 1 and the other dimensions are 0.

[0078] In machine learning, the information processing device 100 executes an error propagation method similar to that of the above-mentioned neural network 140. The information processing device 100 inputs input data corresponding to the mesh model 51 to the graph convolutional neural network 150, evaluates the error between the output and the teacher data 54, and updates the weights of the graph convolutional neural network 150. The error is, for example, the Euclidean distance between two vectors.

[0079] Moreover, the information processing device 100 inputs input data corresponding to the mesh model 52 to the graph convolutional neural network 150, evaluates the error between the output and the teacher data 55, and updates the weights of the graph convolutional neural network 150. Moreover, the information processing device 100 inputs input data corresponding to the mesh model 53 to the graph convolutional neural network 150, evaluates the error between the output and the teacher data 56, and updates the weights of the graph convolutional neural network 150.

[0080] FIG. 11 is a diagram showing an example of stepwise pooling for a mesh model. According to the above-described graph convolutional neural network 150, the fine mesh model 31 is gradually simplified through the pooling layers 152-1 to 152-4. The information processing device 100 obtains the coarse mesh model 32 suitable for the finite element method simulation by extracting the output of an appropriate pooling layer from among the pooling layers 152-1 to 152-4. The information processing device 100 may ignore the class vector that is the final output.

[0081] For example, the information processing device 100 inputs a mesh model 61 to the graph convolutional neural network 150. Then, a mesh model 62 is obtained from the output of the pooling layer 152-1. The mesh model 62 has fewer edges and is simplified compared to the mesh model 61. A mesh model 63 is obtained from the output of the pooling layer 152-2. The mesh model 63 has even fewer edges and is simplified compared to the mesh model 62. A mesh model 64 is obtained from the output of the pooling layer 152-3. The mesh model 64 has even fewer edges and is simplified compared to the mesh model 63. A mesh model 65 is obtained from the output of the pooling layer 152-4. The mesh model 65 has even fewer edges and is simplified compared to the mesh model 64.

[0082] During machine learning to generate the graph convolutional neural network 150, the user may check the outputs of the pooling layers 152-1 to 152-4 and select an appropriate pooling layer. The user may also perform a trial finite element method simulation and select a pooling layer taking into consideration the execution time and accuracy of the rough simulation result 33.

[0083] Next, the convolution operation (mesh convolution operation) and the pooling operation (mesh pooling operation) of the graph convolutional neural network 150 will be described. The mesh convolution operation and the mesh pooling operation are also described in the following documents, for example. Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman and Daniel Cohen-Or, "MeshCNN: A Network with an Edge", ACM Transactions on Graphics, Volume 38, Issue 4, Article No. 90, July 2019.

[0084] FIG. 12 is a diagram illustrating an example of parameters of the edge feature amount. First, the information processing device 100 calculates the feature amount (initial edge feature amount) of each edge included in the input mesh model. Edge 70 is an edge that connects node 71 and node 72. Edge 70 belongs to a triangular face formed by nodes 71, 72, and 73. Edge 70 also belongs to a triangular face formed by nodes 71, 72, and 74. In this way, one edge belongs to two faces.

[0085] The information processing device 100 calculates an edge feature e corresponding to the edge 70. The edge feature e is a five-dimensional vector. The first dimension indicates a dihedral angle that is the angle between the two faces to which the edge 70 belongs. The second dimension indicates an inner angle of a vertex of one of the faces. For example, the second dimension is the inner angle of node 74. The third dimension indicates an inner angle of a vertex of the other face. For example, the third dimension is the inner angle of node 73.

[0086] The fourth dimension is the edge length ratio, which is the ratio of the base to the height of one face, i.e., the ratio between the length of edge 70 and the length of a perpendicular line drawn from a vertex (e.g., node 74) of one face to edge 70. The fifth dimension is the edge length ratio, which is the ratio of the base to the height of the other face, i.e., the ratio between the length of edge 70 and the length of a perpendicular line drawn from a vertex (e.g., node 73) of the other face to edge 70.

[0087] The convolution layers 151-1 to 151-4 perform mesh convolution calculations for each remaining edge to update the edge feature. The mesh convolution calculations use edge feature values ​​of adjacent edges, which are the other four edges included in the two faces to which the edge of interest belongs. Thus, the edge feature values ​​of the five edges are processed as one set.

[0088] In the following, the processing of the convolution layer 151-1 will be described as a representative example. According to the definition of the edge feature above, the convolution layer 151-1 acquires an edge feature e of the edge 70. In addition, the convolution layer 151-1 acquires an edge feature a of the adjacent edge between the nodes 72 and 73, an edge feature b of the adjacent edge between the nodes 71 and 73, an edge feature c of the adjacent edge between the nodes 71 and 74, and an edge feature d of the adjacent edge between the nodes 72 and 74.

[0089] Here, the order of the four adjacent edges is specified as counterclockwise with respect to edge 70. However, since the priority of two faces is not uniquely determined on a mesh model, the order of the four edge features can be two ways: (a, b, c, d) and (c, d, a, b). Therefore, in order to guarantee the invariance of the mesh convolution operation, the convolution layer 151-1 converts the edge features a, b, c, d into edge features e1, e2, e3, e4 as shown in Equation (1). The edge features e1, e2, e3, e4 are symmetrical.

[0090]

number

[0091] The convolution layer 151-1 updates the edge feature e as shown in Equation (2) based on the edge feature e of the edge 70 and the edge features e1, e2, e3, and e4 of the four adjacent edges. k0, k1, k2, k3, and k4 in Equation (2) are kernel coefficients determined by machine learning and indicate weights of the five edge features. Note that the edge features e, e1, e2, e3, and e4 are five-dimensional vectors. The kernel coefficients k0, k1, k2, k3, and k4 may be common among the five dimensions or may be different for each dimension.

[0092]

number

[0093] The above mesh convolution operations can be implemented as matrix operations. For example, the convolution layer 151-1 has a size of N c ×N e Creates a tensor of size N × 5. c is the number of dimensions of each edge feature, and N c = 5. e is the number of edges to be considered in one mesh convolution operation. For example, N e = 1. The final "5" means the total number of edges including the edge of interest. The convolution layer 151-1 holds a kernel matrix in which kernel coefficients are arranged. The convolution layer 151-1 executes GeMM (General Matrix Multiplication) as a dot product between the above tensor and the kernel matrix. By preparing multiple kernel matrices, multiple edge feature amounts can be calculated for each edge.

[0094] FIG. 13 is a diagram showing an example of edge removal by pooling. Each of the pooling layers 152-1 to 152-4 removes a specific number of edges (for example, 150 edges) based on the edge feature amount calculated in the preceding convolutional layer. The process of the pooling layer 152-1 will be described below as a representative example.

[0095] The pooling layer 152-1 calculates the L2 norm of the edge feature for each edge. The L2 norm is the square root of the sum of squares that indicates the Euclidean distance from the zero vector that is the origin of the vector space. The pooling layer 152-1 preferentially selects a specific number of edges in order from the edge with the smaller L2 norm. The edges selected here are edges with low importance. The pooling layer 152-1 removes the selected edge and integrates the nodes at both ends of the edge into one. The position of the node after integration is, for example, the middle of the positions of the nodes at both ends. As a result, two of the four adjacent edges disappear. The pooling layer 152-1 updates the edge feature of the remaining two adjacent edges.

[0096] For example, consider a case where edges 81, 82, and 83 form one face, and edges 81, 84, and 85 form one face. Edge 81 corresponds to the above-mentioned edge 70 and has an edge feature e. Edge 82 has an edge feature a, edge 83 has an edge feature b, edge 84 has an edge feature c, and edge 85 has an edge feature d.

[0097] When the L2 norm of the edge feature amount e is minimum, the pooling layer 152-1 removes the edge 81. Then, the edges 82 and 83 are integrated into the edge 86, and the edges 84 and 85 are integrated into the edge 87. In this way, when one edge with low importance is selected, five edges are integrated into two edges, and two faces disappear.

[0098] The pooling layer 152-1 calculates an edge feature amount p of an edge 86 and an edge feature amount q of an edge 87. The edge feature amount p is an average vector of the edge feature amounts e, a, and b of the edges 81, 82, and 83 that form one face. The edge feature amount q is an average vector of the edge feature amounts e, c, and d of the edges 81, 84, and 85 that form one face.

[0099] The pooling layer 152-1 may select edges to be removed by considering the distribution of nodes or edges in the entire input mesh model. For example, the pooling layer 152-1 detects an area in the input mesh model where the density of nodes or the density of edges exceeds a threshold. The pooling layer 152-1 selects and removes an edge with a minimum L2 norm from the detected area. In this case, the pooling layer 152-1 does not need to remove edges included in low-density areas even if the L2 norm is small.

[0100] Next, the functions and processing procedures of the information processing device 100 will be described. Note that, although the information processing device 100 of the second embodiment executes both machine learning and stress prediction, different information processing devices can execute machine learning and stress prediction. In this case, a learning model is transmitted from one information processing device to the other information processing device.

[0101] FIG. 14 is a block diagram illustrating an example of functions of the information processing device. The information processing device 100 has a mesh storage unit 121, a sample result storage unit 122, and a learning model storage unit 123. These storage units are implemented using, for example, a RAM 102 or a HDD 103. The information processing device 100 also has a graph CNN learning unit 124, an NN learning unit 125, a coarse mesh generation unit 126, a finite element method execution unit 127, and a stress prediction unit 128. These processing units are implemented using, for example, a CPU 101 or a GPU 104, and a program.

[0102] The mesh storage unit 121 stores mesh models. The stored mesh models include samples used for machine learning of the graph convolution neural network 150. The stored mesh models also include samples used for generating teacher data for the neural network 140. The stored mesh models also include a fine mesh model 31 that is a target of stress prediction.

[0103] The sample result storage unit 122 stores the results of the finite element method simulation used for machine learning of the neural network 140. The stored simulation results include fine simulation results generated from samples of the fine mesh model. The learning model storage unit 123 stores the neural network 140 and the graph convolution neural network 150 generated by machine learning.

[0104] The graph CNN learning unit 124 reads samples of mesh models 51, 52, 53, etc. from the mesh storage unit 121 and generates training data. The graph CNN learning unit 124 uses the generated training data to generate a graph convolutional neural network 150 by backpropagation. The graph CNN learning unit 124 stores the generated graph convolutional neural network 150 in the learning model storage unit 123.

[0105] The NN learning unit 125 reads out the simulation results of the samples from the sample result storage unit 122 and generates training data. The NN learning unit 125 uses the generated training data to generate a neural network 140 by the backpropagation method. The NN learning unit 125 stores the generated neural network 140 in the learning model storage unit 123.

[0106] The coarse mesh generation unit 126 reads the fine mesh model 31 from the mesh storage unit 121, and reads the graph convolutional neural network 150 from the learning model storage unit 123. The coarse mesh generation unit 126 generates input data from the fine mesh model 31 and inputs it to the graph convolutional neural network 150. The coarse mesh generation unit 126 extracts the output of the pooling layer 152-2 to generate the coarse mesh model 32.

[0107] For machine learning, the finite element method execution unit 127 reads out a sample of the fine mesh model from the mesh storage unit 121. The finite element method execution unit 127 executes a numerical analysis method as an iterative method on the sample. The finite element method execution unit 127 generates a fine simulation result indicating the calculated stress, and stores it in the sample result storage unit 122.

[0108] Furthermore, for stress prediction, the finite element method execution unit 127 receives the coarse mesh model 32 from the coarse mesh generation unit 126. The finite element method execution unit 127 executes a numerical analysis method as an iterative method on the coarse mesh model 32. The finite element method execution unit 127 generates a coarse simulation result 33 indicating the calculated stress, and transmits it to the stress prediction unit 128.

[0109] The stress prediction unit 128 reads out the neural network 140 from the learning model storage unit 123, and receives the coarse simulation result 33 from the finite element method execution unit 127. The stress prediction unit 128 generates input data using the coarse simulation result 33, inputs the input data to the neural network 140, and generates a fine simulation result 34 indicating a predicted value of stress. The stress prediction unit 128 outputs the fine simulation result 34. The stress prediction unit 128 may store the fine simulation result 34 in the HDD 103, may display it on the display device 111, or may transmit it to another information processing device.

[0110] FIG. 15 is a diagram showing an example of the structure of the mesh data and the training data. The mesh storage unit 121 stores a node table 131 and a face table 132 for each mesh model. The node table 131 includes, for each of a plurality of nodes, a node ID that identifies the node and three-dimensional coordinates that indicate the position of the node. The face table 132 includes, for each of a plurality of faces, a face ID that identifies the face and the node IDs of the three nodes that form the face.

[0111] The graph CNN learning unit 124 generates a training data table 133. The training data table 133 includes a plurality of records each including an edge feature and a class vector of 750 edges. The 750 edge features correspond to explanatory variables and input data. The class vector corresponds to objective variables and teacher data. Each edge feature is a five-dimensional vector including a dihedral angle, an interior angle 1, an interior angle 2, an edge length ratio 1, and an edge length ratio 2. The class vector is a 30-dimensional vector in which only one dimension is "1".

[0112] The NN learning unit 125 generates a training data table 134. The training data table 134 includes a plurality of records each including a corner angle, a mesh size, a fillet radius, a stress ratio, and a fillet stress. The corner angle, the mesh size, the fillet radius, and the stress ratio correspond to explanatory variables and input data. The fillet stress corresponds to objective variables and teacher data. The corner angle, the mesh size, the fillet radius, and the stress ratio are provided by the user as 625 types of simulation conditions. The fillet stress is calculated in the finite element method execution unit 127.

[0113] FIG. 16 is a flowchart showing an example of a machine learning procedure. (S10) The graph CNN learning unit 124 extracts edges from the fine mesh model samples, and calculates edge features of each edge based on the relationship between the edge of interest and four adjacent edges. The graph CNN learning unit 124 generates input data including edge features of a predetermined number of connected edges. The graph CNN learning unit 124 also generates class vectors as teacher data from labels attached to the fine mesh model samples. The graph CNN learning unit 124 generates training data in which the input data and teacher data are associated with each other.

[0114] (S11) The graph CNN learning unit 124 generates a graph convolutional neural network 150 by the backpropagation method using the training data generated in step S10. (S12) The finite element method execution unit 127 obtains a sample of a fine mesh model that satisfies each of the 625 simulation conditions described above, and executes a finite element method simulation to generate a fine simulation result.

[0115] (S13) The NN learning unit 125 extracts a fillet from among the samples of the fine mesh model. The fillet may be specified by a user. The NN learning unit 125 may also detect the fillet by pattern matching.

[0116] (S14) The NN learning unit 125 generates input data indicating parameters of each fillet from the simulation conditions of step S12. The NN learning unit 125 also generates teacher data indicating the fillet stress calculated in step S12. The NN learning unit 125 generates training data in which the input data and the teacher data are associated with each other.

[0117] (S15) The NN learning unit 125 generates the neural network 140 by the backpropagation method using the training data generated in step S14. (S16) The finite element method execution unit 127 stores the graph convolution neural network 150 generated in step S11. The NN learning unit 125 stores the neural network 140 generated in step S15.

[0118] FIG. 17 is a flowchart showing an example of a procedure for the stress simulation. (S20) The coarse mesh generation unit 126 acquires the fine mesh model 31 to be simulated, extracts edges from the fine mesh model 31, and calculates the edge feature amount of each edge from the relationship between the edge of interest and four adjacent edges. The coarse mesh generation unit 126 generates input data including the edge feature amounts of a predetermined number of connected edges.

[0119] (S21) The coarse mesh generation unit 126 inputs input data to the graph convolutional neural network 150. The coarse mesh generation unit 126 extracts the output of the pooling layer 152-2 in the graph convolutional neural network 150, and generates a coarse mesh model 32 that indicates the edges output by the pooling layer 152-2.

[0120] (S22) The finite element method execution unit 127 executes a finite element method simulation on the coarse mesh model 32 to generate a coarse simulation result 33. (S23) The stress prediction unit 128 extracts corners from within the coarse mesh model 32. The corners may be specified by the user. Alternatively, the stress prediction unit 128 may detect the corners by pattern matching.

[0121] (S24) The stress prediction unit 128 generates input data including the corner angle 41 and mesh size 42 indicated by the coarse mesh model 32, the fillet radius 43 indicated by the fine mesh model 31, and the stress ratio 44 indicated by the coarse simulation result 33.

[0122] (S25) The stress prediction unit 128 inputs the input data of step S24 to the neural network 140, and predicts the fillet stress 45. The stress prediction unit 128 generates the micro simulation result 34 including the predicted fillet stress 45.

[0123] (S26) The stress prediction unit 128 outputs the microscopic simulation result 34. For example, the stress prediction unit 128 stores the microscopic simulation result 34 in a non-volatile storage such as the HDD 103. Also, for example, the stress prediction unit 128 displays the microscopic simulation result 34 on the display device 111. Also, for example, the stress prediction unit 128 transmits the microscopic simulation result 34 to another information processing device.

[0124] As described above, the information processing device 100 of the second embodiment generates and outputs the fine simulation result 34 indicating the fillet stress from the fine mesh model 31 indicating the fillet. Thus, information useful for designing industrial parts is provided. Furthermore, the information processing device 100 generates the coarse simulation result 33 from the coarse mesh model 32 corresponding to the fine mesh model 31 by finite element method simulation, and predicts the fine simulation result 34 from the coarse simulation result 33. Thus, the amount of calculation is reduced and the execution time is shortened compared to the case where the fine simulation result 34 is directly generated from the fine mesh model 31 by finite element method simulation.

[0125] Moreover, the information processing device 100 automatically generates the coarse mesh model 32 from the fine mesh model 31 by using the graph convolutional neural network 150. This reduces the burden on the user of creating the coarse mesh model 32 separately from the fine mesh model 31, and improves the efficiency of the structural analysis simulation.

[0126] Moreover, the information processing device 100 generates the coarse mesh model 32 based on the output of an intermediate pooling layer included in the graph convolutional neural network 150. Here, the graph convolutional neural network 150 is trained to gradually remove edges with low importance through multiple pooling layers. Thus, a high-quality coarse mesh model 32 that preserves the essential features of the fine mesh model 31 is generated. As a result, the accuracy of the coarse simulation result 33 and the fine simulation result 34 is improved. [Explanation of symbols]

[0127] 10. Information processing device 11 Storage section 12 Processing section 13,14 Pre-trained models 15,16 Mesh data 17,18 Simulation result data

Claims

1. On the computer, generating, from second mesh data, first mesh data having fewer edges than the second mesh data, using a first trained model that removes some edges from mesh data including a plurality of nodes and edges between the nodes to determine features of the mesh data; executing a simulation using the first mesh data to generate first simulation result data indicating a physical quantity of an object indicated by the first mesh data; inferring, from the first mesh data and the first simulation result data, second simulation result data corresponding to a case where the simulation is performed using the second mesh data, using a second trained model; An inference program that executes the processing.

2. The first trained model is a convolutional neural network including a plurality of pooling layers that gradually reduce the number of edges, and the first mesh data is generated based on an output of one of the plurality of pooling layers. The inference program according to claim 1.

3. the first trained model is a class determination model that determines a class of an object represented by the mesh data, and the first mesh data is generated by extracting intermediate data from the first trained model when determining the class of the second mesh data. The inference program according to claim 1.

4. the second mesh data is fine mesh data that indicates a fillet, and the first mesh data is coarse mesh data that indicates an acute angle that is sharper than the fillet, The second trained model infers a second stress generated in the fillet from a first parameter corresponding to a shape of the acute angle, a second parameter corresponding to a first stress generated in the acute angle, and a third parameter corresponding to a shape of the fillet. The inference program according to claim 1.

5. The computer generating, from second mesh data, first mesh data having fewer edges than the second mesh data, using a first trained model that removes some edges from mesh data including a plurality of nodes and edges between the nodes to determine features of the mesh data; executing a simulation using the first mesh data to generate first simulation result data indicating a physical quantity of an object indicated by the first mesh data; inferring, from the first mesh data and the first simulation result data, second simulation result data corresponding to a case where the simulation is performed using the second mesh data, using a second trained model; Reasoning method.

6. A storage unit that stores a first trained model that determines a feature of mesh data by removing some edges from mesh data including a plurality of nodes and edges between the nodes; a processing unit that uses the first trained model to generate, from second mesh data, first mesh data having fewer edges than the second mesh data, executes a simulation using the first mesh data to generate first simulation result data indicating a physical quantity of an object indicated by the first mesh data, and infers, from the first mesh data and the first simulation result data, second simulation result data corresponding to a case where the simulation is executed using the second mesh data, using a second trained model; An information processing device having the above configuration.

Citation Information

Patent Citations

  • Resolution control system and resolution control program

    JP2005242647A

  • Analytical mesh correction device

    JP2006318232A

  • Mesh generation device for analysis

    JP2011243016A

  • Mesh generation device

    JP2013178599A

  • Application of convolutional neural networks to object meshes

    US20180286120A1