Simulation program, simulation apparatus, and simulation method
By dynamically switching between neural networks and FEM for generating stiffness matrices based on error indices, the simulation program enhances the accuracy and speed of structural analysis simulations.
Patent Information
- Application Number
- JP2023082154
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-18
- Publication Date
- 2024-11-28
AI Technical Summary
In structural analysis simulations, using a neural network to generate a stiffness matrix instead of a numerical calculation method based on FEM leads to decreased analysis accuracy over time due to accumulating errors, which affects the convergence of nonlinear algorithms.
A simulation program that dynamically switches between a faster but less accurate first generation method, such as a neural network, and a slower but more accurate second generation method like FEM, based on an error index, to maintain calculation accuracy.
This approach improves the calculation accuracy of physical quantities by correcting errors in the stiffness matrix, ensuring high-speed and accurate structural analysis simulations.
Smart Images

Figure 2024165727000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a simulation technique. [Background technology]
[0002] A structural analysis simulation is a simulation that calculates the temporal and spatial changes of physical quantities related to an object to which an external force is applied. A simulation program for a structural analysis simulation calculates physical quantities such as displacement and stress by solving partial differential equations using, for example, the finite element method (FEM).
[0003] With respect to FEM, computational windowing techniques are known for finite element mechanical, thermomechanical, electromechanical, and / or magnetomechanical analysis (see, for example, U.S. Pat. No. 5,393,313 and U.S. Pat. No. 5,493,633). Methods of using computer-aided engineering (CAE) analysis to perform numerical simulations of physical behavior of physical domains on a computer (see, for example, U.S. Pat. No. 5,493,313 and U.S. Pat. No. 5,493,633). [Prior art documents] [Patent documents]
[0004] [Patent Document 1] US Patent Application Publication No. 2020 / 0226310 [Patent Document 2] Special Publication No. 2022-517229 [Patent Document 3] US Patent Application Publication No. 2014 / 0343899 [Patent Document 4] JP 2014-225257 A Summary of the Invention [Problem to be solved by the invention]
[0005] In a structural analysis simulation of an object, if a stiffness matrix is generated using a neural network trained by machine learning instead of a numerical calculation method based on FEM, the accuracy of the analysis decreases as the time step progresses. In the following, the numerical calculation method based on FEM may be simply referred to as FEM.
[0006] This problem is not limited to the case where a stiffness matrix is generated by a neural network instead of FEM, but occurs when a stiffness matrix is generated by various generation methods with lower accuracy than FEM.
[0007] In one aspect, the present invention aims to improve the calculation accuracy of a simulation in which physical quantities are calculated while generating a stiffness matrix of an object to be analyzed. [Means for solving the problem]
[0008] In one proposal, the simulation program causes a computer to execute the following processes.
[0009] The computer repeats a first analysis process based on a first generation method for generating a stiffness matrix of the analysis object. The first analysis process includes a process of generating a stiffness matrix of the analysis object by the first generation method, and a process of calculating a physical quantity related to the analysis object using the stiffness matrix generated by the first generation method.
[0010] The computer calculates an index related to an error in the stiffness matrix generated by the first generation method after the first analysis process is repeated, and changes the first generation method to a second generation method having higher accuracy than the first generation method based on the index related to the error.
[0011] The computer repeats a second analysis process based on the second generation method. The second analysis process includes a process of generating a stiffness matrix of the analysis object by the second generation method, and a process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the second generation method. Effect of the Invention
[0012] According to one aspect, it is possible to improve the calculation accuracy of a simulation in which a physical quantity is calculated while generating a stiffness matrix of an analysis target. [Brief description of the drawings]
[0013] [Figure 1] FIG. 13 is a diagram showing a change in stress over time in a structural analysis simulation. [Diagram 2] FIG. 2 is a functional configuration diagram of a simulation device according to an embodiment. [Diagram 3] 1 is a flowchart of a first simulation process. [Figure 4] FIG. 2 is a functional configuration diagram showing a specific example of a simulation device. [Diagram 5] FIG. 4 is a diagram showing a change in stress over time in a structural analysis simulation according to an embodiment. [Figure 6] FIG. 13 is a diagram showing a first time change of the L1 norm. [Figure 7] FIG. 13 is a diagram showing a time change of the L2 norm. [Figure 8] FIG. 13 is a diagram showing a change in maximum value norm over time. [Figure 9] FIG. 13 is a diagram showing a second time change of the L1 norm. [Figure 10A] 13 is a flowchart (part 1) of a second simulation process. [Figure 10B] 13 is a flowchart (part 2) of the second simulation process. [Figure 10C] 13 is a flowchart (part 3) of the second simulation process. [Figure 11] FIG. 2 is a hardware configuration diagram of an information processing device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0014] Hereinafter, the embodiments will be described in detail with reference to the drawings.
[0015] In nonlinear dynamic analysis of structural analysis simulations of objects, an analytical method that combines a nonlinear algorithm such as the Newton-Raphson method with the time marching method is sometimes used.
[0016] In such an analysis method, an approximation calculation step is repeated one or more times for each time step. In each approximation calculation step, a new stiffness matrix is generated, and an approximate solution of the physical quantities is calculated by solving simultaneous linear equations with the generated stiffness matrix as a coefficient matrix. The approximation calculation step is sometimes called an iteration.
[0017] In nonlinear dynamic analysis using a supercomputer, when a stiffness matrix is generated using a numerical calculation method based on FEM at each approximation calculation step, the time required to calculate the stiffness matrix often takes much longer than the time required to calculate the solution to the simultaneous linear equations. As a result, the time required to generate the stiffness matrix takes up the majority of the execution time of the structural analysis simulation.
[0018] On the other hand, when a stiffness matrix is generated by a neural network trained by machine learning, the time required to generate the stiffness matrix is shorter than that required by FEM. For example, in a simulation of the analysis of the earth's crust structure during an earthquake, the calculation time for the stiffness matrix using a neural network is about one-third of the calculation time for the stiffness matrix using FEM.
[0019] However, it is difficult to train a neural network to obtain highly accurate analysis results over all time steps of a structural analysis simulation. Therefore, when the generation of the stiffness matrix by the neural network is repeated over a certain number of time steps, a large error accumulates. The accumulated stiffness matrix error seriously affects the convergence of the nonlinear algorithm that calculates the physical quantities, and the accuracy of the calculated physical quantities decreases.
[0020] In other words, a neural network can generate a stiffness matrix faster than FEM, but the accuracy of the stiffness matrix generated by a neural network is lower than that of the stiffness matrix generated by FEM. Therefore, the analysis accuracy of a structural analysis simulation using a neural network is lower than that of a structural analysis simulation using FEM.
[0021] Figure 1 shows an example of the time evolution of stress in a simulation of the structure analysis of the earth's crust during an earthquake. In this example, the earth's crust in a specific region is divided into multiple elements in the FEM. Each element is represented by a hexahedron with eight nodes.
[0022] The horizontal axis represents time steps, and the vertical axis represents stress in a given direction at a specific node of a specific element. The "E" on the vertical axis represents the power of 10. The polygonal line 101 represents the change in stress over time when the stiffness matrix is generated by FEM. The polygonal line 102 represents the change in stress over time when the stiffness matrix is generated by a neural network.
[0023] Comparing the polygonal lines 101 and 102, it can be seen that as the time step progresses, the accuracy of the stress calculation when using a neural network decreases and the difference in stress gradually increases. For example, the difference in stress between the polygonal lines 101 and 102 at time step 5 is 0.00015, and the difference in stress between the polygonal lines 101 and 102 at time step 10 is 0.0055.
[0024] 2 shows an example of the functional configuration of a simulation device according to an embodiment. A simulation device 201 in FIG.
[0025] Fig. 3 is a flowchart showing an example of a first simulation process performed by the simulation device 201 of Fig. 2. First, the first analysis unit 211-1 repeats a first analysis process based on a first generation method for generating a stiffness matrix of an object to be analyzed (step 301). The first analysis process includes a process for generating a stiffness matrix of the object to be analyzed by the first generation method, and a process for calculating a physical quantity related to the object to be analyzed by using the stiffness matrix generated by the first generation method.
[0026] Next, the change unit 212 calculates an index related to an error of the stiffness matrix generated by the first generation method after the first analysis process is repeated (step 302). Then, the change unit 212 changes the first generation method to a second generation method having higher accuracy than the first generation method based on the index related to the error (step 303).
[0027] Next, the second analysis unit 211-2 repeats the second analysis process based on the second generation method (step 304). The second analysis process includes a process of generating a stiffness matrix of the analysis object by the second generation method, and a process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the second generation method.
[0028] According to the simulation device 201 in FIG. 2, it is possible to improve the calculation accuracy of a simulation in which physical quantities are calculated while generating a stiffness matrix of an analysis target.
[0029] Fig. 4 shows a specific example of the simulation device 201 in Fig. 2. The simulation device 401 in Fig. 4 includes a division unit 411, a generation unit 412-1, a generation unit 412-2, a calculation unit 413-1, a calculation unit 413-2, a generation unit 414-1, a generation unit 414-2, a control unit 415, an output unit 416, and a storage unit 417.
[0030] The generation unit 412-1 and the calculation unit 413-1 correspond to the first analysis unit 211-1 in Fig. 2, and the generation unit 412-2 and the calculation unit 413-2 correspond to the second analysis unit 211-2 in Fig. 2. The generation unit 414-1, the generation unit 414-2, and the control unit 415 correspond to the modification unit 212 in Fig. 2.
[0031] The simulation device 401 performs a structural analysis simulation of an object in CAE, structural design, engineering design, etc. The object to be subjected to the structural analysis simulation may be a structure such as a road, a bridge, or a building, the earth's crust, a fluid, etc., or an industrial product.
[0032] In a structural analysis simulation, an approximation calculation step is repeated one or more times for each time step. In each approximation calculation step, a new stiffness matrix is generated, and physical quantities such as displacement and stress are calculated by solving simultaneous linear equations with the generated stiffness matrix as a coefficient matrix. Displacement or stress corresponds to a physical quantity related to an object.
[0033] The storage unit 417 stores object information 421 that represents the shape of an object. The division unit 411 divides the object represented by the object information 421 into a plurality of elements, generates element information 422 that represents each of the plurality of elements, and stores the element information 422 in the storage unit 417. Each element has a plurality of nodes, and the element information 422 of each element includes identification information and position information of each node of the element. The shape of each element may be a polyhedron. The plurality of elements generated by the division unit 411 is an example of a plurality of elements that represent an analysis target.
[0034] In each approximation calculation step of each time step, the generation unit 412-1 generates an element stiffness matrix of each element by the first generation method using the element information 422 of the element. In generating the element stiffness matrix, the physical quantity of each node calculated in the immediately preceding time step or the physical quantity of each node calculated in the immediately preceding approximation calculation step of the same time step is used.
[0035] As the first generation method, for example, a generation method that generates an element stiffness matrix having lower accuracy than FEM at a higher speed than FEM is used. As the first generation method, for example, a generation method that generates an element stiffness matrix using a trained machine learning model 423 stored in the storage unit 417. As the machine learning model 423, a neural network, a random forest, or the like is used.
[0036] The machine learning model 423 uses the physical quantities of each of the multiple nodes of an element as an input, infers an element stiffness matrix of the element, and outputs the inferred element stiffness matrix. The machine learning model 423 may infer element stiffness matrices of two or more elements simultaneously. By using the machine learning model 423, the element stiffness matrix can be generated at high speed.
[0037] The machine learning model 423 is generated by training a pre-learning machine learning model by supervised machine learning. For example, the physical quantity of each node of each of the multiple elements in the first time step of the structural analysis simulation can be used as the input training data. In addition, the element stiffness matrix of each element used as the correct answer label is generated from the same training data by, for example, a second generation method described later.
[0038] The calculation unit 413-1 generates a stiffness matrix of the entire object using the element stiffness matrix of each element generated by the first generation method in each approximation calculation step of each time step. Then, the calculation unit 413-1 calculates the physical quantity of each node of each element using the stiffness matrix of the entire object.
[0039] The output unit 416 outputs the physical quantities calculated in the final approximation calculation step of each time step as the analysis results for that time step.
[0040] The processing performed by generation unit 412-1 and calculation unit 413-1 corresponds to a first analysis processing based on a first generation method.
[0041] In each approximation calculation step of each time step, the generation unit 412-2 generates an element stiffness matrix of each element by the second generation method using the element information 422 of the element. In generating the element stiffness matrix, the physical quantity of each node calculated in the immediately preceding time step or the physical quantity of each node calculated in the immediately preceding approximation calculation step of the same time step is used.
[0042] As the second generation method, for example, a generation method is used that generates an element stiffness matrix having higher accuracy than the first generation method, and takes a longer time than the first generation method. The second generation method is, for example, FEM. The FEM may be a numerical calculation method such as the F-bar method. By using the FEM, a more accurate element stiffness matrix can be generated.
[0043] The calculation unit 413-2 generates a stiffness matrix of the entire object using the element stiffness matrix of each element generated by the second generation method in each approximation calculation step of each time step. Then, the calculation unit 413-2 calculates the physical quantity of each node of each element using the stiffness matrix of the entire object.
[0044] The output unit 416 outputs the physical quantities calculated in the final approximation calculation step of each time step as the analysis results for that time step.
[0045] The processing performed by the generation unit 412-2 and the calculation unit 413-2 corresponds to a second analysis processing based on a second generation method.
[0046] The control unit 415 evaluates the error of the stiffness matrix to be generated next by the first generation method in the time step of error evaluation after the first analysis process is repeated by the generation unit 412-1 and the calculation unit 413-1. Then, the control unit 415 determines whether or not to change the first generation method to the second generation method based on the evaluation result. The time step of the error evaluation may be set at a rate of once every fixed number of time steps.
[0047] In a time step of error evaluation, the generating unit 414-1 uses one or more elements among the elements represented by the element information 422 as evaluation target elements. Then, the generating unit 414-1 generates an element stiffness matrix M1 of each evaluation target element by a first generation method in the same manner as the generating unit 412-1. The element stiffness matrix M1 is an example of a first element stiffness matrix.
[0048] Moreover, the generating unit 414-2 generates an element stiffness matrix M2 of each evaluation target element by the second generating method in the same manner as the generating unit 412-2. The element stiffness matrix M2 is an example of a second element stiffness matrix.
[0049] The control unit 415 calculates an index ER relating to the error of the stiffness matrix using the element stiffness matrix M1 and element stiffness matrix M2 of each evaluation target element. As the index ER, for example, a statistical value of the mean absolute error (MAE) of the element stiffness matrix M1 and element stiffness matrix M2 is used. The control unit 415 calculates, for example, E(M1,M2) representing the MAE of the element stiffness matrix M1 and element stiffness matrix M2 by the following formula.
[0050] E(M1,M2) =(1 / N 2 )*Σ i Σ j |M2(i,j)-M1(i,j)| (1)
[0051] N represents the number of rows and columns of the element stiffness matrix M1. The number of rows and columns of the element stiffness matrix M2 is also N. Σ i represents the summation for i = 1 to N, and Σ j represents the sum for j = 1 to N. For example, if each element has eight nodes and each node has three degrees of freedom, then N = 3*8 = 24.
[0052] M1(i,j) represents the matrix element in row i and column j of the element stiffness matrix M1, and M2(i,j) represents the matrix element in row i and column j of the element stiffness matrix M2. |M2(i,j)-M1(i,j)| represents the absolute value of M2(i,j)-M1(i,j).
[0053] As the statistical value of E(M1,M2), the average value, minimum value, maximum value, etc. of E(M1,M2) of all the elements to be evaluated is used. The number of elements to be evaluated may be about one hundredth to one tenth of the total number of elements representing the object. Instead of E(M1,M2) in formula (1), the average error, the root mean square error, etc. may be used.
[0054] The control unit 415 evaluates an error in the stiffness matrix to be generated next by the first generation method by comparing the index ER with a threshold value TH1. The threshold value TH1 may be specified by a user or calculated based on experimental results.
[0055] If the index ER is greater than the threshold TH1, the control unit 415 determines that an error due to the first generation method has accumulated, and changes the first generation method to a second generation method that is more accurate than the first generation method. In this case, the first analysis process by the generation unit 412-1 and the calculation unit 413-1 ends, and the second analysis process by the generation unit 412-2 and the calculation unit 413-2 starts.
[0056] When errors in the stiffness matrix accumulate due to repeated execution of the first analysis process, the method for generating the stiffness matrix can be switched to the second generation method, thereby reducing errors in the stiffness matrix generated at each time step after the switch. This improves the convergence of the nonlinear algorithm and increases the accuracy of the calculated physical quantities.
[0057] If the index ER is equal to or smaller than the threshold TH1, the control unit 415 determines that the error due to the first generation method has not accumulated, and continues to apply the first generation method. In this case, the first analysis process by the generation unit 412-1 and the calculation unit 413-1 is repeated until the time step for the next error evaluation is reached.
[0058] In addition, by calculating the index ER using the element stiffness matrix M1 and the element stiffness matrix M2 of the element to be evaluated, it is possible to appropriately evaluate whether or not an error due to the first generation method has accumulated with a small amount of calculation.
[0059] When the first generation method is changed to the second generation method, the generation unit 412-1 and the calculation unit 413-1 repeat the second analysis process for X time steps. Then, after the second analysis process is repeated X times, the control unit 415 changes the second generation method back to the first generation method. In this case, the second analysis process by the generation unit 412-2 and the calculation unit 413-2 ends, and the first analysis process by the generation unit 412-1 and the calculation unit 413-1 is resumed. X is an example of a predetermined number of times.
[0060] By switching the stiffness matrix generation method to the first generation method after the second analysis process has been repeated X times, the structural analysis simulation can be performed at high speed using the stiffness matrix with reduced error in the subsequent time steps. Therefore, the structural analysis simulation can be performed at high speed while suppressing a decrease in calculation accuracy.
[0061] X may be specified by a user or may be determined based on the index ER. When X is determined based on the index ER, the control unit 415 calculates X at a time step of error evaluation, for example, by the following formula.
[0062] X=int((|ER-TH2| / TH2)+1)*10 (2)
[0063] int() is a function that converts the value in parentheses into an integer, and represents the largest integer that is equal to or smaller than the value in parentheses. The threshold value TH2 may be specified by a user or calculated based on experimental results. According to formula (2), the larger the absolute value of the difference between ER and TH2, the larger the value of X.
[0064] By increasing the number of iterations X of the second analysis process when ER is large, the error in the stiffness matrix is corrected by the second generation method over a long period of time, improving the calculation accuracy at each time step after the first analysis process is resumed.
[0065] Fig. 5 shows an example of the time change of stress in the structural analysis simulation of the embodiment. The analysis target of the structural analysis simulation in Fig. 5 is the earth's crust as in Fig. 1, and the method of dividing the earth's crust is also the same as in Fig. 1. In the simulation result in Fig. 5, a broken line 501 is added to the simulation result in Fig. 1.
[0066] A broken line 501 indicates the change in stress over time calculated by the simulation device 401 in Fig. 4. As the first generation method, a method of generating a stiffness matrix using a neural network is used, and as the second generation method, FEM is used. In time steps 1 to 5, the stiffness matrix is generated using a neural network, and in time steps 6 to 10, the stiffness matrix is generated by FEM.
[0067] In the case of polygonal line 501, by switching the stiffness matrix generation method to FEM in time step 6, the error is dynamically corrected in time steps 6 to 10, and the calculated stress gradually approaches that of polygonal line 101. Therefore, a more accurate analysis result can be obtained than in the case of polygonal line 102, for which the stiffness matrix generation method is not switched.
[0068] Next, an example of the change over time in the displacement error in the structural analysis simulation of FIG. 5 will be described with reference to FIGS.
[0069] In the following description, simulation S0 represents a structural analysis simulation in which a stiffness matrix is generated by FEM at every time step. Simulation S1 represents a structural analysis simulation in which a stiffness matrix is generated by using a neural network at every time step. Simulation S2 represents a structural analysis simulation in which a stiffness matrix is generated by switching between the neural network and FEM using the simulation device 401 in FIG. 4.
[0070] The displacement error in simulation S1 represents the difference between the displacement calculated in simulation S1 and the displacement calculated in simulation S0. This error is calculated using the displacement of each of the multiple nodes calculated in simulation S1 and the displacement of each of the multiple nodes calculated in simulation S0.
[0071] The displacement error in simulation S2 represents the difference between the displacement calculated in simulation S2 and the displacement calculated in simulation S0. This error is calculated using the displacement of each of the multiple nodes calculated in simulation S2 and the displacement of each of the multiple nodes calculated in simulation S0.
[0072] In the examples of FIGS. 6 to 9, the L1 norm, the L2 norm, or the maximum norm is used as the displacement error in the simulation S1 or the simulation S2.
[0073] 6 shows an example of a first time change of the L1 norm. The horizontal axis represents the time step, and the vertical axis represents the L1 norm. A broken line 601 shows the time change of the L1 norm in the simulation S1. A broken line 602 shows the time change of the L1 norm in the simulation S2.
[0074] In simulation S2, the method of generating the stiffness matrix is switched every time five time steps have passed, and the stiffness matrix is generated using a neural network in time steps 1 to 5 and time steps 11 to 15. In contrast, the stiffness matrix is generated by FEM in time steps 6 to 10 and time steps 16 to 20. Periods 611 and 612 represent periods during which FEM was used as the method of generating the stiffness matrix.
[0075] The L1 norm indicated by the polygonal line 601 increases monotonically from time step 1 to time step 14, and diverges from time step 15 onwards.
[0076] The L1 norm indicated by the broken line 602 increases in time step 1 to time step 5, where the neural network was used, similarly to the broken line 601. Next, the L1 norm decreases monotonically in time step 6 to time step 10, where the FEM was used, and it is understood that the error is corrected by the FEM.
[0077] Next, in time steps 11 to 15, where the neural network is used, the L1 norm increases monotonically again. Next, in time steps 16 to 20, where the FEM is used, the L1 norm continues to increase, and it can be seen that five time steps are not enough time to correct the error.
[0078] 7 shows an example of the change over time of the L2 norm. The horizontal axis represents the time step, and the vertical axis represents the L2 norm. A broken line 701 shows the change over time of the L2 norm in simulation S1. A broken line 702 shows the change over time of the L2 norm in simulation S2. A period 711 and a period 712 represent the periods in which FEM was used as a method for generating the stiffness matrix in simulation S2.
[0079] 6, the L2 norm indicated by the polygonal line 702 decreases monotonically from time step 6 to time step 10, where FEM is used, and it is clear that the error is corrected by FEM. On the other hand, the L2 norm continues to increase from time step 16 to time step 20, where FEM is used, and it is clear that five time steps are insufficient time to correct the error.
[0080] Fig. 8 shows an example of the change over time of the maximum value norm. The horizontal axis represents the time step, and the vertical axis represents the maximum value norm. A broken line 801 shows the change over time of the maximum value norm in simulation S1. A broken line 802 shows the change over time of the maximum value norm in simulation S2. A period 811 and a period 812 represent the periods during which FEM was used as a method for generating the stiffness matrix in simulation S2.
[0081] The maximum value norm indicated by the polygonal line 801 increases monotonically from time step 1 to time step 14, and diverges from time step 15 onwards.
[0082] The maximum value norm indicated by line 802 increases in time step 1 to time step 5, where the neural network was used, similar to line 801. Next, the maximum value norm decreases monotonically in time step 6 to time step 10, where FEM was used, and it is understood that the error is corrected by FEM.
[0083] Next, in time steps 11 to 15, where the neural network is used, the maximum norm increases monotonically again. Next, in time steps 16 to 20, where the FEM is used, the maximum norm decreases once and then remains constant.
[0084] Fig. 9 shows a second example of time variation of the L1 norm. The horizontal axis represents the time step, and the vertical axis represents the L1 norm. The broken line 601 shows the time variation of the L1 norm in the simulation S1, as in Fig. 6. The broken line 901 shows the time variation of the L1 norm in the simulation S2.
[0085] 9, after the method of generating the stiffness matrix of simulation S2 is changed to FEM in time step 16, FEM is continuously used for 25 time steps. Periods 911 and 913 represent periods during which the neural network is used as the method of generating the stiffness matrix. Periods 912 and 914 represent periods during which FEM is used as the method of generating the stiffness matrix.
[0086] The L1 norm indicated by the broken line 901 changes in the same manner as the broken line 602 in Fig. 6 from time step 1 to time step 15. Next, from time step 16 to time step 40 in the period 914 in which the FEM was used, the L1 norm increases once and then decreases monotonically, indicating that the error is corrected by the FEM. It can be said that the period 914, which includes 25 time steps, is sufficient time to correct the error.
[0087] The simulation device 401 can perform not only a structural analysis simulation of an object, but also a simulation of other analysis targets such as an electric field, a magnetic field, etc.
[0088] 10A to 10C are flowcharts showing an example of the second simulation process performed by the simulation device 401 of FIG.
[0089] First, the control unit 415 sets the control variable t indicating the time step to 1, which indicates the first time step (step 1001), and sets the control variable C indicating the method of generating the stiffness matrix to G1, which indicates the first generation method (step 1002).
[0090] Next, the control unit 415 checks the value of C (step 1003). If C=G1 (step 1003, YES), the control unit 415 selects the first generation method as the method for generating the stiffness matrix (step 1004).
[0091] Next, the generator 412-1 generates an element stiffness matrix of each element by the first generation method using the element information 422 of that element (step 1005).
[0092] Next, the calculation unit 413-1 generates a stiffness matrix for the entire object using the element stiffness matrix of each element generated by the first generation method (step 1006).Then, the calculation unit 413-1 calculates the physical quantity of each node of each element using the stiffness matrix of the entire object (step 1007).
[0093] Next, the calculation unit 413-1 checks whether the approximation calculation for calculating the physical quantity has converged (step 1008). If the approximation calculation has not converged (step 1008, NO), the simulation device 401 repeats the processes from step 1005 onwards as the next approximation calculation step.
[0094] If the approximation calculation has converged (step 1008, YES), the output unit 416 outputs the calculated physical quantity as the analysis result at the current time step (step 1009). The processing of steps 1004 to 1009 corresponds to a first analysis processing based on the first generation method.
[0095] Next, the control unit 415 checks whether t indicates the last time step of the structural analysis simulation (step 1010). If t does not indicate the last time step (step 1010, NO), the control unit 415 increments t by 1 (step 1011). Then, the control unit 415 checks whether t indicates the time step of error evaluation (step 1012).
[0096] If t indicates the time step of error evaluation (step 1012, YES), the generation unit 414-1 generates an element stiffness matrix M1 of each evaluation target element by the first generation method using the element information 422 of the evaluation target element (step 1013).
[0097] Next, the generating unit 414-2 generates an element stiffness matrix M2 of each evaluation target element by the second generating method using the element information 422 of the evaluation target element (step 1014).
[0098] Next, the control unit 415 calculates an index ER using the element stiffness matrix M1 and the element stiffness matrix M2 of each evaluation target element (step 1015), and compares the index ER with a threshold value TH1 (step 1016).
[0099] If the index ER is greater than the threshold value TH1 (step 1016, YES), the control unit 415 sets G2, which indicates the second generation method, to C (step 1017). Then, the simulation device 401 repeats the processes from step 1003 onwards.
[0100] If t does not indicate the time step of error evaluation (step 1012, NO), or if the index ER is equal to or smaller than the threshold value TH1 (step 1016, NO), the simulation device 401 repeats the processes from step 1003 onwards.
[0101] If t indicates the last time step (step 1010, YES), the simulation device 401 ends the process.
[0102] If C=G2 (step 1003, NO), the control unit 415 selects the second generation method as the method for generating the stiffness matrix (step 1018).
[0103] Next, the generating unit 412-2 generates an element stiffness matrix of each element by the second generating method using the element information 422 of that element (step 1019).
[0104] Next, the calculation unit 413-2 generates a stiffness matrix for the entire object using the element stiffness matrix of each element generated by the second generation method (step 1020).Then, the calculation unit 413-2 calculates the physical quantity of each node of each element using the stiffness matrix of the entire object (step 1021).
[0105] Next, the calculation unit 413-2 checks whether the approximation calculation for calculating the physical quantity has converged (step 1022). If the approximation calculation has not converged (step 1022, NO), the simulation device 401 repeats the processes from step 1019 onwards as the next approximation calculation step.
[0106] If the approximation calculation has converged (step 1022, YES), the output unit 416 outputs the calculated physical quantity as the analysis result at the current time step (step 1023). The processing of steps 1018 to 1023 corresponds to a second analysis processing based on the second generation method.
[0107] Next, the control unit 415 checks whether t indicates the last time step (step 1024). If t does not indicate the last time step (step 1024, NO), the control unit 415 increments t by 1 (step 1025). Then, the control unit 415 checks whether the second analysis process based on the second generation method has been repeated for X time steps (step 1026).
[0108] When the second analysis process has been repeated X times (step 1026, YES), the control unit 415 sets G1 to C (step 1027). Then, the simulation device 401 repeats the processes from step 1003 onwards.
[0109] If the second analysis process has not been repeated X times (step 1026, NO), the simulation device 401 repeats the processes from step 1003 onwards.
[0110] If t indicates the last time step (step 1024, YES), the simulation device 401 ends the process.
[0111] The configurations of the simulation device 201 in FIG. 2 and the simulation device 401 in FIG. 4 are merely examples, and some of the components may be omitted or changed depending on the application or conditions of the simulation device.
[0112] The flowcharts of FIG. 3 and FIGS. 10A to 10C are merely examples, and some of the processes may be omitted or changed depending on the configuration or conditions of the simulation device.
[0113] The time changes of stress shown in Figures 1 and 5 are merely examples, and the time changes of stress change depending on the analysis object. The time changes of displacement error shown in Figures 6 to 9 are merely examples, and the time changes of displacement error change depending on the analysis object.
[0114] The formulas (1) and (2) are merely examples, and the simulation device 401 may calculate the indices ER and X using other formulas.
[0115] FIG. 11 shows an example of the hardware configuration of an information processing device (computer) used as the simulation device 201 in FIG. 2 and the simulation device 401 in FIG.
[0116] 11 includes a CPU (Central Processing Unit) 1101, a memory 1102, an input device 1103, and an output device 1104. The information processing device further includes an auxiliary storage device 1105, a media drive device 1106, a network connection device 1107, and a GPU (Graphics Processing Unit) 1108. These components are hardware, and are connected to each other via a bus 1109.
[0117] The memory 1102 is, for example, a semiconductor memory such as a Read Only Memory (ROM) or a Random Access Memory (RAM), and stores programs and data used in processing. The memory 1102 may operate as the storage unit 417 in FIG.
[0118] 2 by executing a program using the memory 1102. The CPU 1101 (processor) also performs a process of calculating the physical quantity of the first analysis unit 211-1 by executing a program using the memory 1102.
[0119] 4 by executing a program using memory 1102. CPU 1101 may also operate as generation unit 412-1 and generation unit 414-1 by executing a program using memory 1102.
[0120] The input device 1103 is, for example, a keyboard, a pointing device, etc., and is used to input instructions or information from a user or operator. The output device 1104 is, for example, a display device, a printer, etc., and is used to output inquiries or instructions to a user or operator and processing results. The output device 1104 may operate as the output unit 416 in Fig. 4, and the processing results may be analysis results.
[0121] The auxiliary storage device 1105 is, for example, a magnetic disk device, an optical disk device, a magneto-optical disk device, a tape device, or the like. The auxiliary storage device 1105 may be a hard disk drive or an SSD (Solid State Drive). The information processing device can store programs and data in the auxiliary storage device 1105 and use them by loading them into the memory 1102. The auxiliary storage device 1105 may operate as the storage unit 417 in FIG. 4.
[0122] The medium drive device 1106 drives the portable recording medium 1110 and accesses the recorded contents. The portable recording medium 1110 is a memory device, a flexible disk, an optical disk, a magneto-optical disk, etc. The portable recording medium 1110 may be a CD-ROM (Compact Disk Read Only Memory), a DVD (Digital Versatile Disk), a USB (Universal Serial Bus) memory, etc. A user or operator can store programs and data in the portable recording medium 1110 and load them into the memory 1102 for use.
[0123] Thus, the computer-readable recording medium that stores the program and data used in the processing is a physical (non-transitory) recording medium such as the memory 1102, the auxiliary storage device 1105, or the portable recording medium 1110.
[0124] The network connection device 1107 is a communication interface circuit that is connected to a communication network such as a WAN (Wide Area Network) or a LAN (Local Area Network) and performs data conversion associated with communication. The information processing device can receive programs and data from an external device via the network connection device 1107 and load them into the memory 1102 for use. The network connection device 1107 may operate as the output unit 416 in FIG. 4.
[0125] The GPU 1108 is a calculation processing circuit that performs information processing. The GPU 1108 may operate as the generation unit 412-1 and the generation unit 414-1 in FIG.
[0126] It is not necessary for the information processing device to include all of the components in Fig. 11, and some of the components may be omitted or changed depending on the purpose or conditions of the information processing device. For example, if an interface with a user or operator is not required, the input device 1103 and the output device 1104 may be omitted. If the portable recording medium 1110 or a communication network is not used, the medium drive device 1106 or the network connection device 1107 may be omitted.
[0127] Although the disclosed embodiments and their advantages have been described in detail, it will be appreciated that those skilled in the art may make various modifications, additions and omissions therein without departing from the scope of the invention as clearly set forth in the claims.
[0128] The following notes are further disclosed regarding the embodiment described with reference to FIGS. (Appendix 1) repeating a first analysis process based on a first generation method for generating a stiffness matrix of an analysis object; calculating an index relating to an error of a stiffness matrix generated by the first generation method after the first analysis process is repeated; changing the first generation method to a second generation method having a higher accuracy than the first generation method based on the error indicator; repeating a second analysis process based on the second generation method; The process is executed by a computer, The first analysis process includes: A process of generating a stiffness matrix of the analysis object by the first generation method; A process of calculating a physical quantity related to the analysis object using the stiffness matrix generated by the first generation method; Including, The second analysis process includes: A process of generating a stiffness matrix of the analysis object by the second generation method; A process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the second generation method; A simulation program comprising: (Appendix 2) The process of calculating the error-related indicator includes: a process of using any one of the plurality of elements representing the analysis target as an evaluation target element and generating a first element stiffness matrix of the evaluation target element by the first generation method; generating a second element stiffness matrix of the evaluation target element by the second generation method; a process of calculating an index related to the error using the first element stiffness matrix and the second element stiffness matrix; 2. The simulation program according to claim 1, (Appendix 3) The process of repeating the second analysis process includes a process of repeating the second analysis process a predetermined number of times, The simulation program according to claim 1, further comprising causing the computer to execute a process of changing the second generation method to the first generation method and repeating the first analysis process based on the first generation method after the second analysis process has been repeated the predetermined number of times. (Appendix 4) 4. The simulation program according to claim 3, wherein the predetermined number of times is determined based on an index related to the error. (Appendix 5) The simulation program according to any one of claims 1 to 4, wherein the first generation method is a generation method using a machine learning model, and the second generation method is a numerical calculation method based on the finite element method. (Appendix 6) a first analysis unit that repeats a first analysis process based on a first generation method for generating a stiffness matrix of an analysis object; a change unit that calculates an index related to an error of a stiffness matrix generated by the first generation method after the first analysis process is repeated, and changes the first generation method to a second generation method having higher accuracy than the first generation method based on the index related to the error; a second analysis unit that repeats a second analysis process based on the second generation method after the first generation method is changed to the second generation method; Equipped with The first analysis process includes: A process of generating a stiffness matrix of the analysis object by the first generation method; A process of calculating a physical quantity related to the analysis object using the stiffness matrix generated by the first generation method; Including, The second analysis process includes: A process of generating a stiffness matrix of the analysis object by the second generation method; A process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the second generation method; A simulation device comprising: (Appendix 7) the change unit uses any one of a plurality of elements representing the analysis target as an evaluation target element, generates a first element stiffness matrix of the evaluation target element by the first generation method, generates a second element stiffness matrix of the evaluation target element by the second generation method, and calculates an index related to the error by using the first element stiffness matrix and the second element stiffness matrix. (Appendix 8) The second analysis unit repeats the second analysis process a predetermined number of times, the change unit changes the second generation method to the first generation method after the second analysis process has been repeated the predetermined number of times; 7. The simulation device according to claim 6, wherein the first analysis unit repeats the first analysis process based on the first generation method after the second generation method is changed to the first generation method. (Appendix 9) 9. The simulation device according to claim 8, wherein the predetermined number of times is determined based on an index related to the error. (Appendix 10) The simulation device according to any one of claims 6 to 9, wherein the first generation method is a generation method using a machine learning model, and the second generation method is a numerical calculation method based on a finite element method. (Appendix 11) repeating a first analysis process based on a first generation method for generating a stiffness matrix of an analysis object; calculating an index relating to an error of a stiffness matrix generated by the first generation method after the first analysis process is repeated; changing the first generation method to a second generation method having a higher accuracy than the first generation method based on the error indicator; repeating a second analysis process based on the second generation method; The computer executes the process, The first analysis process includes: A process of generating a stiffness matrix of the analysis object by the first generation method; A process of calculating a physical quantity related to the analysis object using the stiffness matrix generated by the first generation method; Including, The second analysis process includes: A process of generating a stiffness matrix of the analysis object by the second generation method; A process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the second generation method; A simulation method comprising: (Appendix 12) The process of calculating the error-related indicator includes: a process of using any one of the plurality of elements representing the analysis target as an evaluation target element and generating a first element stiffness matrix of the evaluation target element by the first generation method; generating a second element stiffness matrix of the evaluation target element by the second generation method; a process of calculating an index related to the error using the first element stiffness matrix and the second element stiffness matrix; 12. The simulation method according to claim 11, comprising: (Appendix 13) The process of repeating the second analysis process includes a process of repeating the second analysis process a predetermined number of times, The simulation method of claim 11, further comprising the step of: changing the second generation method to the first generation method after the second analysis process has been repeated the predetermined number of times, and repeating the first analysis process based on the first generation method, wherein the computer further executes the step of: (Appendix 14) 14. The simulation method according to claim 13, wherein the predetermined number of times is determined based on an index related to the error. (Appendix 15) The simulation method according to any one of claims 11 to 14, wherein the first generation method is a generation method using a machine learning model, and the second generation method is a numerical calculation method based on the finite element method. [Explanation of symbols]
[0129] 101, 102, 501, 601, 602, 701, 702, 801, 802, 901 Line 201, 401 Simulation device 211-1 1st Analysis Department 211-2 2nd Analysis Department 212 Changes 411 Split section 412-1, 412-2, 414-1, 414-2 generation section 413-1, 413-2 Calculation section 415 Control Unit 416 Output section 417 Storage section 421 Object information 422 Element Information 423 Machine Learning Models 611, 612, 711, 712, 811, 812, 911~914 period 1101 CPU 1102 Memory 1103 Input Device 1104 Output device 1105 Auxiliary storage device 1106 Media drive unit 1107 Network connection device 1108 GPU 1109 Bus 1110 Portable recording media
Claims
1. repeating a first analysis process based on a first generation method for generating a stiffness matrix of an analysis object; calculating an index relating to an error of a stiffness matrix generated by the first generation method after the first analysis process is repeated; changing the first generation method to a second generation method having a higher accuracy than the first generation method based on the error indicator; repeating the second analysis process based on the second generation method; The process is executed by a computer, The first analysis process includes: A process of generating a stiffness matrix of the analysis object by the first generation method; a process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the first generation method; Including, The second analysis process includes: generating a stiffness matrix of the analysis target by the second generation method; a process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the second generation method; A simulation program comprising:
2. The process of calculating the error-related indicator includes: a process of using any one of a plurality of elements representing the analysis target as an evaluation target element, and generating a first element stiffness matrix of the evaluation target element by the first generation method; generating a second element stiffness matrix of the evaluation target element by the second generation method; a process of calculating an index related to the error using the first element stiffness matrix and the second element stiffness matrix; 2. The simulation program according to claim 1, further comprising:
3. repeating the second analysis process includes repeating the second analysis process a predetermined number of times; 2. The simulation program according to claim 1, further comprising causing the computer to execute a process of changing the second generation method to the first generation method and repeating the first analysis process based on the first generation method after the second analysis process has been repeated the predetermined number of times.
4. 4. The simulation program according to claim 3, wherein the predetermined number of times is determined based on an index relating to the error.
5. The simulation program according to any one of claims 1 to 4, characterized in that the first generation method is a generation method using a machine learning model, and the second generation method is a numerical calculation method based on a finite element method.
6. a first analysis unit that repeats a first analysis process based on a first generation method for generating a stiffness matrix of an analysis object; a change unit that calculates an index related to an error of a stiffness matrix generated by the first generation method after the first analysis process is repeated, and changes the first generation method to a second generation method having higher accuracy than the first generation method based on the index related to the error; a second analysis unit that repeats a second analysis process based on the second generation method after the first generation method is changed to the second generation method; Equipped with The first analysis process includes: A process of generating a stiffness matrix of the analysis object by the first generation method; a process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the first generation method; Including, The second analysis process includes: generating a stiffness matrix of the analysis target by the second generation method; a process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the second generation method; A simulation device comprising:
7. repeating a first analysis process based on a first generation method for generating a stiffness matrix of an analysis object; calculating an index relating to an error of a stiffness matrix generated by the first generation method after the first analysis process is repeated; changing the first generation method to a second generation method having a higher accuracy than the first generation method based on the error indicator; repeating the second analysis process based on the second generation method; The computer executes the process, The first analysis process includes: A process of generating a stiffness matrix of the analysis object by the first generation method; a process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the first generation method; Including, The second analysis process includes: generating a stiffness matrix of the analysis target by the second generation method; a process of calculating a physical quantity related to the analysis object by using the stiffness matrix generated by the second generation method; A simulation method comprising:
Citation Information
Patent Citations
Method and system for dividing undirected graph expressing matrix to be used for cae by using bidirectional level set
JP2014225257A
Systems and methods for computational windowing
JP2022517229A
Methods and systems for using bi-directional level sets to partition an undirected graph representing a matrix to be used in CAE
US20140343899A1
Systems and Methods for Computational Windowing
US20200226310A1
