Generation device, generation method and program
The generation of a surrogate model for homogenization constitutive law addresses computational and formulation challenges in multiscale analysis, enabling efficient and accurate analysis of inelastic materials in commercial software.
Patent Information
- Application Number
- JP2023084859
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-23
- Publication Date
- 2024-12-05
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Conventional multiscale analysis methods for nonlinear problems face challenges in computational cost and formulation difficulties, particularly in commercial software, and lack support for macro analysis in inelastic materials.
A generation device and method that generates a surrogate model of the homogenization constitutive law based on multiple numerical material tests, enabling separated multiscale analysis by representing the relationship between macro stress, macro strain, and macro cumulative strain.
Enables efficient implementation of separated multiscale analysis in commercial software, expanding the scope of multi-scale analysis of composite materials, particularly for inelastic bodies, with improved accuracy and versatility.
Smart Images

Figure 2024168308000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a generating device for generating a surrogate model. [Background technology]
[0002] In recent years, composite materials have been used in various structures from the viewpoints of safety and reducing environmental load. Since the macroscopic mechanical properties of composite materials depend on their microstructures, simulations using multiscale analysis based on homogenization theory are used to predict the mechanical behavior of such structures (e.g., Non-Patent Documents 1 to 4). [Prior art documents] [Non-patent literature]
[0003] [Non-Patent Document 1] D. Huang et al., Comput. Methods Appl. Mech. Eng., 2020 [Non-Patent Document 2] Y.Fish & Y.Yu, Int. J. Num. Mech. Eng., 2022 [Non-Patent Document 3] F. Fritzen et al., Eur. J. Mech. A / Solids, 2018 [Non-Patent Document 4] F.Ghavamian & A.simone, Comput. Methods Appl. Mech. Eng., 2019 Summary of the Invention [Problem to be solved by the invention]
[0004] The core of multi-scale analysis lies in localization, which solves micro-boundary value problems using conditions determined according to the state of the structure, and homogenization, which calculates macro-stresses and macro-mechanical properties from the solutions of the micro-boundary value problems. Conventionally proposed multi-scale analysis methods for nonlinear problems can be broadly divided into two types depending on the methods of localization and homogenization.
[0005] One is a method called the coupled type (FE2 type), which is a method in which a microanalysis is actually performed at each calculation point in a macroanalysis, and is excellent in that it can accurately obtain solutions to multi-scale problems.The challenge with this method is the calculation cost, and there is a problem that it is unrealistic to use the coupled type to calculate an actual structure for a nonlinear problem. The other is a method called the decoupled type, which uses a homogenization constitutive law that directly calculates macrostress from macrostrain and macro-internal variables, and is superior in that it allows macroanalysis at a significantly lower calculation cost than the coupled type. However, there is a problem in that it is difficult to formulate the homogenization constitutive law (mathematical model). As described above, conventional multiscale analysis methods have their drawbacks, and in recent years, research has been conducted into data-driven multiscale analysis that uses data science techniques to overcome these challenges.
[0006] One method of data-driven multi-scale analysis is to perform multi-scale analysis by reducing the overall calculation cost by reducing the dimension of the micro boundary value problem in the conventional coupled analysis using principal component analysis or the like (hereinafter, for convenience, this method will be referred to as the "dimensionality reduction type"). However, this dimensionality reduction type requires solving the micro problem at each calculation point of the macro calculation, just like the conventional coupled analysis, and therefore has the problem that it cannot be implemented in commercial software that is widely used for general mechanical calculations in practice.
[0007] Another method of data-driven multi-scale analysis is to generate a surrogate model of the homogenization constitutive law in the separation type and directly calculate the macro-stress from the information on the macro-displacement (hereinafter, this method will be referred to as the "surrogate homogenization type" for convenience). With this method, it is not necessary to perform a micro-analysis when solving the macro-boundary value problem, and it can be implemented in general commercial software as a user subroutine. However, most existing research on this method has focused on elastic bodies, and the response of surrogate homogenized models of inelastic materials has been limited to verifying the material response at the material point level, meaning that it is not suitable for macroscopic analysis except for very simple problems.
[0008] The present invention has been made in consideration of such problems, and focuses on the surrogate homogenization type, with one objective being to realize the generation of a surrogate model of a homogenization constitutive law that enables the realization of a separation-type multi-scale analysis. [Means for solving the problem]
[0009] According to a first aspect of the present invention, a generation device for generating a proxy model of a homogenized constitutive law that enables a decoupled multi-scale analysis includes a generation unit that generates a proxy model that represents the relationship between macro-stress and macro-accumulated strain based on the results of multiple numerical material tests performed on a test piece. According to a second aspect of the present invention, a method for generating a surrogate model of a homogenized constitutive law that enables a decoupled multi-scale analysis includes generating a surrogate model that represents the relationship between macro-stress and macro-accumulated strain based on the results of multiple numerical material tests performed on a test piece. According to a third aspect of the present invention, a program for causing a computer to generate a proxy model of a homogenized constitutive law that enables a decoupled multi-scale analysis causes the computer to function as a generation unit that generates a proxy model that represents the relationship between macroscopic strain and macroscopic accumulated strain and macroscopic stress based on the results of multiple numerical material tests performed on a test piece. Effect of the Invention
[0010] According to the present invention, it is possible to generate a surrogate model of a homogenization constitutive law that enables realization of a decoupled multi-scale analysis. Moreover, it is possible to implement the model in commercial software, which significantly expands the scope of practical multi-scale analysis of composite materials. [Brief description of the drawings]
[0011] [Figure 1] FIG. 2 is a diagram showing an example of a functional configuration of an information processing apparatus. [Diagram 2] 11 is a flowchart showing an example of a flow of processing performed by an information processing device. [Diagram 3] An illustration of macro cumulative strain. [Figure 4] FIG. 1 is an explanatory diagram of surrogate model generation. [Diagram 5] FIG. 1 is an explanatory diagram of surrogate model generation. [Figure 6] FIG. 13 is a diagram showing an example of an experimental result. [Figure 7] FIG. 13 is a diagram showing an example of an experimental result. [Figure 8] FIG. 13 is a diagram showing an example of an experimental result. [Figure 9] FIG. 13 is a diagram showing an example of an experimental result. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0012] Hereinafter, an example of an embodiment of the present invention will be described with reference to the drawings. The components described in the present embodiment are merely examples and are not intended to limit the scope of the present invention.
[0013] In this embodiment, as described above, a method for generating a surrogate model of a homogenization constitutive law that can realize a series of calculations in a decoupled multi-scale analysis will be described.
[0014] [Embodiment] FIG. 1 is a diagram showing an example of the configuration of an information processing device 1 according to the present embodiment. The information processing device 1 includes, for example, a macro structure acquisition unit 110, a homogenization analysis unit 120, a macro structure analysis unit 130, and a localization analysis unit 140 as functional units.
[0015] These may be functional units (functional blocks) possessed by, for example, a control unit (processing unit) of the information processing device 1, and the control unit may be configured to have processing circuits such as, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), and an FPGA (Field Programmable Gate Array). Furthermore, the information processing device 1 of the present embodiment may include a processing program stored in a storage unit (not shown). The processing program may be executed to realize the functions of the respective functional units. The information processing device 1 may also be referred to as a computing device or the like.
[0016] The macrostructure acquisition unit 110 acquires macrostructure information of a material to be analyzed (hereinafter, referred to as "analysis target material") via an input unit (not shown) etc. The macrostructure information may include information on the size of the material etc. As an example, the macrostructure acquisition unit 110 acquires macrostructure information (composite material macrostructure information) of a composite material (for example, a cantilever beam, which will be described later) as the material to be analyzed.
[0017] The homogenization analysis unit 120 is a functional unit that performs homogenization analysis based on the macrostructure information acquired by the macrostructure acquisition unit 110, and has, for example, a microstructure numerical model generation unit 121, a numerical material testing unit 123, and a representative model generation unit 125 as functional units.
[0018] The microstructure numerical model generating unit 121 discretizes the minimum periodic structure of the analysis target material from the macrostructure information acquired by the macrostructure acquiring unit 110, for example, using the finite element method (FEM), and generally assumes its periodicity. Specifically, for example, it sets (generates) a numerical specimen (hereinafter referred to as a "test piece") that is a numerical analysis model of a microstructure (hereinafter referred to as a "representative volume element") that represents heterogeneity. The "test piece" may also be referred to as a "unit cell."
[0019] The numerical material test unit 123 performs a numerical material test on a test piece (numerical specimen) set by the microstructure numerical model generation 121. Since the process of the numerical material test is sometimes called homogenization, this functional unit may be called a homogenization unit or a homogenization processing unit. The numerical material test may be a test in which the FEM model of the test piece is regarded as a numerical specimen in a multi-scale analysis based on the homogenization method, and response data of macroscopic strain and macroscopic stress is obtained. Details of this will be described later.
[0020] The surrogate model generation unit 125 generates (estimates) a surrogate model (hereinafter referred to as a "surrogate model" or a "surrogate homogenization model") which is a model that represents the relationship between macro-stress and macro-accumulated strain and represents the homogenization constitutive law based on the results of the numerical material test by the numerical material test unit 123.
[0021] The macrostructure analysis unit 130 analyzes the macrostructure made of the analysis target material based on the macrostructure information acquired by the macrostructure acquisition unit 110 and the proxy model generated by the proxy model generation unit 125. For example, it acquires the load-displacement relationship at the free end of a cantilever beam, which will be described later, and the macrostress-macrostrain relationship in a region of interest.
[0022] The localization analysis unit 140 performs localization analysis of the material to be analyzed based on the results of the macrostructure analysis by the macrostructure analysis unit 130. For example, a numerical material test is performed again using the deformation history at an arbitrary point in a cantilever beam, which will be described later, as input data to obtain the stress distribution (maximum principal stress distribution and minimum principal stress distribution) in the test piece (unit cell).
[0023] <Processing> FIG. 2 is a flowchart showing an example of the flow of processing performed by the control unit of the information processing device 1 in this embodiment. First, the macro structure acquisition unit 110 performs a macro configuration acquisition process to acquire macro structure information of the analysis target material based on an input from an input unit (not shown) (S110). The information processing device 1 may obtain the macro structure information by receiving it from an external device via a communication unit (not shown).
[0024] Next, the microstructure numerical model generation unit 121 performs a microstructure numerical model generation process (S120). Specifically, for example, a test piece to be used in a numerical material test is set based on the macrostructure information acquired in the step of S110.
[0025] Next, the homogenization analysis unit 120 performs a macro strain setting process for the set test piece (S130). Specifically, a macro strain pattern (hereinafter referred to as a "load pattern") to be applied to the test piece as a condition in the numerical material test is set. This setting may be made based on an input by a user who uses the information processing device 1, or may be automatically set by the homogenization analysis unit 120.
[0026] Next, the numerical material test unit 123 performs a numerical material test process based on the load pattern set in step S130 (S140), as will be described in detail later.
[0027] Next, the homogenization analysis unit 120 performs a macro accumulated strain calculation process (S150).
[0028] In this embodiment, a variable called "macro cumulative strain" is introduced as a history parameter that depends on the history of the macro strain. The macro cumulative strain is a variable newly introduced by the inventors of the present application in order to characterize the macro loading state that reflects the deformation history of the material, and can be formulated as shown in the following equation (1).
number
[0029] Here, "tilde ε acc " denotes macro cumulative strain, and "tilde ε" denotes macro strain. Also, the dotted "tilde ε" on the right side indicates the time derivative of the macroscopic strain, that is, the macroscopic strain rate.
[0030] Additionally, “H” is the Heaviside function given by the following equation (2).
number
[0031] "Tilde ε min " and "tilde ε max " are the minimum and maximum values of macroscopic strain in the past, and are expressed as the following equations (3) and (4), respectively.
number
number
[0032] Here, by time discretizing equation (1), the macro cumulative strain can be calculated as shown in the following equation (5).
number
[0033] The subscript “j” indicates the macro cumulative strain component, and the subscript “n” indicates time. Also, the "Δ tilde ε" on the right side is the time step size "Δt = t n+1 -t n " indicates the increment in macroscopic strain between
[0034] The macro accumulated strain calculated in this manner can be regarded as a value calculated by accumulating the time integral of the macro strain rate when the current value of the macro strain is greater than the maximum value of the macro strain previously given as a condition, or less than the minimum value of the macro strain previously given as a condition.
[0035] One of the reasons for introducing the macro cumulative distortion in generating the surrogate model of the present invention is that by taking into account the past history of the macro distortion in the process of generating the surrogate model (the process of learning), it is possible to improve the accuracy of interpolation of behavior that depends on the past history of the macro distortion (improving the regression accuracy in learning).
[0036] FIG. 3 is a diagram for explaining the relationship between the macro strain history (corresponding to the load pattern described above) which is the history of the macro strain, and the macro accumulated strain history which is the history of the macro accumulated strain. Here, an example of the history of macro-accumulated strain calculated according to formula (5) for each macro-strain component (e.g., ε11, ε22, 2ε12) when a monotonic loading test is performed on a specimen is shown. In the graphs shown at the top and bottom of the figure, the horizontal axis indicates time t, and the vertical axis indicates macro-strain and macro-accumulated strain, respectively. The results show that under monotonic loading, the macro-accumulated strain behaves as if it records the maximum / minimum values for each component of the macro-strain.
[0037] By introducing the macro-accumulated strain calculated in this manner as a variable, it becomes possible to generate a surrogate model that takes into account the influence of the macro-strain accumulated due to the load.
[0038] Next, the homogenization analysis unit 120 determines whether or not the collection of data for generating the surrogate model (the collection of data for optimizing the surrogate model parameters) has been completed (S160). If it is determined that data collection has not ended (S160: NO), the homogenization analysis unit 120 returns the process to step S130, and then the homogenization analysis unit 120 resets the load pattern (S130).
[0039] Note that a plurality of load patterns may be set at one time in step S130. If S160: NO, the process may return to step S140.
[0040] If it is determined that data collection has ended (S160: YES), the proxy model generation unit 125 performs a proxy model parameter optimization process (S170).
[0041] Here, a method for generating a surrogate model (a method for optimizing parameters of the surrogate model) will be described in detail. As mentioned above, the decoupled multi-scale analysis method requires a homogenization constitutive law, but it can be difficult to formulate it. Therefore, in this embodiment, teacher data (learning data) is generated by the method described below, and a surrogate model (surrogate homogenization model) is generated that approximates the unlearned parts by, for example, RBF (Radial Basis Function) interpolation.
[0042] In this embodiment, the proxy model is defined, for example, as in the following expressions (6) to (9).
number
number
number
number
[0043] Here, the tilde σ denotes macroscopic stress. RBF interpolation is a method to globally interpolate known input and output data using RBF, where the input data is represented by "f" and the response data by "φ", which is a vector-valued function (a function with three elements for two-dimensional problems and six elements for three-dimensional problems). "N c " indicates the number of data points (total number), and "i" (in parentheses as a subscript) indicates the data point number. "ψ" is an RBF whose independent variables are the Euclidean norm between an arbitrary input and a known data point (i), and "ω" is a weighting coefficient vector. "N i " indicates the dimension of the input data "f".
[0044] A data point indicates any one of the input data "f" from the data set used for interpolation (teaching data, described in detail later). More specifically, the input data "f" is regarded as a discrete point existing in the space spanned by "f" of the entire teaching data, and this is expressed as a data point. For example, in RBF interpolation, the coordinates of the data point are "f", and the Euclidean distance between two points is calculated to perform global interpolation.
[0045] A characteristic feature of this embodiment is that the input data "f" contains the aforementioned macro cumulative distortion in addition to the macro distortion. In this embodiment, for example, the input data “f” is expressed as a 12-dimensional vector (N i = 12), and the response data "φ" is a six-dimensional vector consisting of six components of the macro Cauchy stress.
[0046] It is necessary to determine the specific form of the RBF and the weighting coefficient vector "ω" so that the response of this surrogate model will be an appropriate value. Therefore, in this embodiment, based on the result of the numerical material test by the numerical material test unit 123, the surrogate model generation unit 125 generates an explanatory variable matrix (explanatory variable data) expressed, for example, by the following formula (10) and a response variable matrix (response variable data) expressed, for example, by the following formula (11), which are used as teacher data in generating a surrogate model (machine learning).
number
number
[0047] In the formulas (10) and (11), "n c " indicates the number of data points as mentioned above. The data enclosed in {} on the right side indicates the data for each load pattern (hereinafter referred to as a "case" for convenience), and the subscripts "a, b,..., z" (in parentheses) indicate the case number (hereinafter referred to as the "case number"). In addition, the subscripts "1, 2, . . . , n s " (in brackets) indicates the analysis step in the load pattern.
[0048] In the numerical material test, for example, a set macro strain history (load pattern) is applied to the test piece, and a simulation is performed to calculate the micro stress distribution in the test piece. In this case, it is preferable to calculate the history of the micro stress distribution by applying the macro strain set stepwise to the test piece. This is because the present invention assumes that, for example, a nonlinear (elastic-plastic) material is to be analyzed, and from the viewpoint of the response accuracy of the proxy model, it is essential that the response history during loading or unloading is included in the teacher data. In addition, from the viewpoint of the calculation accuracy of the tangent modulus required when solving the macro analysis by the implicit method, for example, the history data is very important.
[0049] The problem to be solved in the numerical material test can be the governing equations, micro displacement, and periodic boundary conditions expressed by equations (12) to (14), respectively, and the calculation can be performed using the finite element method.
number
number
number
[0050] Macro stress is obtained by volume averaging the micro stress distribution obtained by solving the governing equation. By performing this at each analysis step, historical data of the macro stress-strain relationship can be obtained. In addition, by calculating the macro cumulative strain at each analysis step from the obtained historical data as described above, the explanatory variable matrix and objective variable matrix for one numerical material test case can be obtained. This type of numerical material test is repeated while changing the macro strain history (load pattern) given to the test specimen, for example, so that there is no overlap (corresponding to the above case numbers "a, b, . . . , z"). The explanatory variable matrix and response variable matrix obtained in each case are arranged in order to generate the explanatory variable matrix and response variable matrix as shown above to be used as training data.
[0051] By using the teacher data generated in this manner, the weighting coefficient vector “ω” can be calculated as follows:
number
number
number
number
number
number
[0052] Here, “η” is a regularization parameter (smoothing parameter). In consideration of the possibility of overlearning, in this embodiment, a regularization parameter “η” for L2 regularization is introduced.
[0053] Here, there are various functional forms for the radial basis function "ψ", and it is necessary to select the optimal functional form depending on the training data generated as described above. Therefore, in this embodiment, the radial basis function “ψ” is set as the discrimination parameter “n ψ Assume that the following function is case-specific:
number
[0054] Here, "β" and "η" in the above formula are "n ψ ” is a hyperparameter, just like “n ψ As with ", it is necessary to use the optimal value depending on the training data. For convenience, these parameters may be collectively referred to as "surrogate model parameters."
[0055] However, since it is difficult to theoretically determine these surrogate model parameters, in this embodiment, these surrogate model parameters are optimized using, for example, Differential Evolution (DE). However, the present invention is not limited to this, and a genetic algorithm or the like may also be used.
[0056] In this case, for the sake of calculation efficiency, for example, the surrogate model parameters “β” and “η” are set to real numbers “a β ","a η " and "β=10 aβ ", η=10 aη" and this may be optimized.
[0057] As an example of the optimization, cross validation (CV) is used, which divides the training data and verifies the reproduction accuracy of the training data. However, this is merely an example and is not limiting.
[0058] In addition, the loss function to be minimized can be, for example, a function as shown in the following equations (22) and (23), and the surrogate model parameters are optimized so that the difference between the response of the surrogate model and the response of the training data is minimized.
number
number
[0059] Identification parameter "n ψ Once " is determined, the radial basis function is uniquely determined. Also, the real number "a β ","a η Once " is determined, the parameters "β" and "η" are determined. This determines the weighting coefficient vector “ω” mentioned above, and generates a surrogate model (trained model).
[0060] After step S170, the macro structure analysis unit 130 performs a macro structure analysis process (S180). Specifically, for example, the macro structure of the analysis target material is analyzed based on the macro structure information acquired in step S110 and the proxy model generated in step S170. For example, the load-displacement relationship at the free end of the cantilever beam described later and the macro stress-macro strain relationship in the target region are acquired.
[0061] Next, the localization analysis unit 140 performs a localization analysis process (S190). Specifically, a localized analysis of the material to be analyzed is performed based on the results of the macroscopic structure analysis in S180. For example, when a cantilever beam, which will be described later, is analyzed as a macroscopic structure, the macroscopic strain and macroscopic accumulated strain at the point where the generated stress is large are extracted from the analysis results, and a numerical material test is performed using these as input data to calculate the stress distribution (maximum principal stress distribution and minimum principal stress distribution) inside the test piece corresponding to the macroscopic point.
[0062] Next, the control unit of the information processing device 1 determines whether or not to end the process (S200). If it is determined that the processing should be continued (S200: NO), the control unit may, for example, return the processing to step S120, and the microstructure numerical model generation unit 121 may perform the microstructure numerical model generation processing, set a test piece (e.g., a different type of test piece) again, and generate a representative model. On the other hand, if it is determined that the process should be ended (S200: YES), the control unit ends the process.
[0063] In addition, if it is determined in step S200 that the processing should be continued (S200: NO), the control unit returns the processing to step S110, and the macrostructure acquisition unit 110 performs the macrostructure acquisition processing to acquire macrostructure information (e.g., macrostructure information of another material to be analyzed) again, and generate a proxy model.
[0064] 4 and 5 are explanatory diagrams of the generation of a surrogate model in this embodiment. As shown in Figure 4, a test specimen is set based on the acquired macro-structural information and a numerical material test is performed. In the figure, the macro strain corresponding to each case "a, b, . . . , z" of the numerical material test and the macro cumulative strain calculated according to the above-mentioned formula (5) are combined as follows: (a) Combination of macro strain and macro accumulated strain in case a (b) Combination of macro strain and macro accumulated strain in case b · · · (z) Combination of macro strain and macro accumulated strain in case z An example is shown. Also shown to the right are the macrostresses for each of these cases.
[0065] As shown in Fig. 5, an explanatory variable matrix is generated from the combination of macro strain and macro accumulated strain in each case obtained as described above, and a response variable matrix is generated from the macro stress in each case obtained as described above. That is, in the above process, a plurality of load patterns are set for a test piece (one type of test piece), and a numerical material test is performed using these as described above, thereby generating the explanatory variable matrix shown in formula (10) and the response variable matrix shown in formula (11). Then, the explanatory variable matrix and the response variable matrix generated in this way are used as teacher data to optimize the above-mentioned surrogate model parameters, thereby generating a surrogate model.
[0066] <Experimental results, etc.> FIG. 6 is a diagram showing the macrostructure and microstructure of the composite material to be analyzed. As shown in FIG. 6(a), the macrostructure to be analyzed is a cantilever beam with a length of 50 mm and a width of 20 mm, and a forced displacement is applied to its free end in the vertical direction perpendicular to the longitudinal direction. The graph in Figure 6(b) shows the time variation of the forced displacement applied. The forced displacement is applied until the vertical displacement increases from 0 to its maximum, and then unloaded until it returns to 0. In this calculation example, in order to reduce calculation costs, the macro analysis is performed as a two-dimensional problem under plane strain conditions.
[0067] As for the microstructure, on the other hand, a rectangular reinforcement is embedded in the center of a unit cell of elastic-plastic material, as shown in Figure 6(c). When the matrix is viewed on the y1-y2 plane, the length of each side is 10 x 10 for the matrix, and 6 x 4 for the reinforcement. Since the reinforcement is rectangular, the matrix is an orthotropic material on a macroscale. The volume fraction of the matrix and the reinforcement is 76:24. Regarding the micro-material constitutive law, the matrix is defined by an associated von Mises model using the nonlinear Voce isotropic hardening law, and the reinforcement is treated as a linear elastic body.
[0068] Figure 6(d) shows the material parameters of the matrix and reinforcement. In this example, the microstructure has a large gap between the elastic modulus of the matrix and that of the reinforcement. In addition, the analytical model of the unit cell was created using a mesh of 200 20-node hexahedral elements, and the reinforcement and base material elements were rigidly connected by sharing nodes.
[0069] By conducting a numerical material test on this unit cell, the aforementioned teacher data for constructing a surrogate homogenization model was generated. In this example, 1100 cases of numerical material tests were conducted while changing the applied macro strain history (load pattern), and one of the results that well represents the macro mechanical properties of this non-homogeneous material was graphed (Fig. 7(a)).
[0070] Figure 7(a) shows the results of a numerical material test in which a 1% macro shear strain was applied in the y1-y2 plane direction and then unloaded. From this graph, it can be seen that the macro stress when the macro strain is applied to the maximum is different from the macro stress when it yields again during unloading.
[0071] This type of behavior is represented by a model called composite hardening, which combines an isotropic hardening model and a kinematic hardening model. However, as mentioned above, in the microstructure of this example, an isotropic hardening model is used for the base material, and neither the kinematic hardening nor composite hardening models are used. Nevertheless, the reason why behavior similar to composite hardening appeared in the macroscopic stress-strain relationship is thought to be that a large gap in the physical properties of the matrix and reinforcement materials caused deformation inside the unit cell to be concentrated in a certain area, resulting in localized plastic strain.
[0072] Furthermore, Figure 7(a) shows the deformation state of the unit cell just before unloading, just after re-yielding, and at the end of unloading, along with the equivalent plastic strain distribution. Just before unloading, it can be seen that plastic deformation is concentrated at the corners of the unit cell. The distribution trend remains the same after unloading, and it can be seen that plastic deformation develops further between re-yielding and unloading. In this way, as deformation is concentrated in some elements within the unit cell, complex hardening behavior appears in the macroscopic stress-strain relationship.
[0073] Next, a surrogate model was constructed using this macroscopic stress-strain relationship, as shown in Figure 7(b). 14,410 pieces of response data from the results of 1,100 numerical material tests were used to optimize the model parameters. Figure 7(b) shows the parameter values. Using the teacher data as interpolation data and performing RBF interpolation, a surrogate homogenized model with a response surface as shown in the figure was obtained.
[0074] Of the three response surfaces, the left surface is the response surface for the horizontal component of macroscopic stress, the center surface is the response surface for the vertical component of macroscopic stress, and the right surface is the response surface for the shear component of macroscopic stress. It can be seen that the horizontal and vertical response surfaces show different shapes, reflecting the orthotropic nature of the microstructure.
[0075] Using this surrogate homogenization model, a macroscopic stress analysis of the cantilever beam mentioned above was carried out (see Fig. 6(e)). A mesh created with 1,000 eight-node quadrilateral elements was used as the analytical model of the target beam, and calculations were carried out using the finite element method. The macroscopic governing equation was used to calculate the macroscopic stress at each calculation point from the surrogate homogenization model, using the equation shown in Fig. 7(c). Note that although the tangent coefficient can also be analytically derived from the surrogate homogenization model, in this example it was calculated using a numerical tangent in order to avoid calculation instability due to vibration of the response surface.
[0076] In addition, to compare with the macroscopic stress analysis using the surrogate homogenization model, direct analyses were performed using meshes of 4,000 elements (40 unit cells), 25,000 elements (250 unit cells), and 100,000 elements (1,000 unit cells). This direct analysis is a method of performing an analysis using a mesh that directly reflects the microstructure of a non-homogeneous material, and differs from a multi-scale analysis in that the unit cells have a finite size.
[0077] In the macroscopic calculations in multi-scale analysis, the size of the unit cell is approximated to 0 based on the homogenization theory, so the results of the direct analysis asymptotically approach the results of the multi-scale analysis as the size of the unit cell becomes smaller.
[0078] Figure 8(a) shows the load-displacement relationship at the free end of the beam obtained by each analysis. The solid line with markers is the result obtained by the surrogate homogenization model, and shows that the reaction force at the peak displacement point is the largest among the comparison objects. Next, in descending order of reaction force at the peak displacement point, are direct analysis (100,000 elements), direct analysis (25,000 elements), and direct analysis (4,000 elements).
[0079] The solid error curve is the error curve between the surrogate homogenization model and direct analysis (4,000 elements), and it has the largest error among the comparisons at the peak displacement point. Next, in descending order of error at the peak displacement point, are direct analysis (25,000 elements) and direct analysis (100,000 elements).
[0080] From this graph, it can be confirmed that the load-displacement relationships obtained by the four types of analysis show the same tendency.In addition, at the maximum displacement, the error in the reaction force of the surrogate homogenized model is about 17% compared to that of 4,000 elements, about 10% compared to that of 25,000 elements, and about 7% compared to that of 100,000 elements, which confirms that the error between the direct analysis and the surrogate homogenized model decreases as the number of elements (number of unit cells) increases.
[0081] This result suggests that the solution obtained using the surrogate homogenization model is the convergent solution when the number of elements in the direct analysis is increased to infinity, and as mentioned above, it is consistent with the principle derived from homogenization theory that the results of the direct analysis asymptotically approach the multi-scale analysis as the unit cell size becomes smaller. Therefore, based on the above results, the solution obtained using the surrogate homogenization model is considered to be a valid solution.
[0082] Figure 8(b) shows the macro-stress-macro-strain relationship in the region of interest of the beam (region common to each analysis) obtained by each analysis. The solid line with markers is the result obtained by the surrogate homogenization model, and the solid line without markers is the result obtained by the direct analysis. As shown in the graph, the results of the surrogate homogenization model and the direct analysis show similar trends. However, since the direct analysis value is not a strict macro-stress, although it can be said that the results of the surrogate homogenization model and the direct analysis show similar trends qualitatively, a quantitative comparison is necessary.
[0083] For quantitative comparison, the stress within the unit cell was compared. Specifically, a localization analysis was performed in which the macroscopic strain history shown in the graph was applied to the unit cell as a boundary condition, and the stress within the unit cell was calculated in an analysis using a surrogate homogenization model. The stress distribution within the target region obtained by the localization analysis was compared with that obtained by a direct analysis, and close results were obtained between the surrogate homogenization model + localization analysis and the direct analysis (100,000 elements).
[0084] Figure 9(b) shows the maximum principal stress distribution and minimum principal stress distribution at maximum loading. When the relative error of the maximum value of the maximum principal stress and the relative error of the minimum value of the minimum principal stress at this time were calculated, the results shown in the table in Figure 9(a) were obtained. The error for the maximum principal stress was 1.99%, and the error for the minimum principal stress was 4.78%. This error also shows that similar results are obtained between the surrogate homogenization model + localization analysis and the direct analysis (100,000 elements).
[0085] From the above results, it was found that the method presented in this embodiment was able to generate a surrogate homogenized model capable of implementing decoupled multi-scale of elastic-plastic materials.
[0086] In this embodiment, the proxy model is generated using RBF interpolation, but the present invention is not limited to this, and a proxy model may be generated using a method other than RBF interpolation. For example, a surrogate model may be generated by applying a neural network model and optimizing the weights of the neural network model using the same method as above. In this case, similar to the above, a cost (e.g., an L2 regularization term) may be added to the loss function of the neural network model to perform optimization.
[0087] <Effects of the embodiment> In this embodiment, an information processing device 1 (e.g., an example of a generating device) includes a numerical material testing unit 123 that performs multiple numerical material tests on a test piece, and a surrogate model generating unit 125 (e.g., an example of a generating unit) that generates a surrogate model that represents the relationship between macro stress and macro strain and macro accumulated strain (a combination of macro strain and macro accumulated strain) based on the results of the numerical material testing. As a result, a homogenized constitutive law surrogate model that enables decoupled multiscale analysis is generated. In addition, the use of macro-accumulated strain improves the accuracy and versatility of the surrogate model generation.
[0088] In this case, the proxy model generating unit 125 may generate a proxy model based on the macro stress obtained by performing multiple numerical material tests on the test piece, and the macro strain and macro accumulated strain stored as conditions for the numerical material test. This makes it possible to appropriately generate a proxy model based on the macro stress obtained by performing multiple numerical material tests on the test piece, and the macro strain and macro accumulated strain stored as conditions for the numerical material test.
[0089] In this case, the proxy model generating unit 125 may calculate the macro accumulated strain based on the macro strain that is a condition of the numerical material test. This makes it possible to simply and appropriately calculate the macro cumulative strain based on the macro strain used as a condition for the numerical material test.
[0090] In this embodiment, the proxy model generating unit 125 generates a proxy model based on the results of a numerical material test performed under conditions of a plurality of different macro strain patterns (load patterns). This makes it possible to appropriately generate a surrogate model based on the results of a numerical material test performed under conditions of a plurality of different macroscopic strain patterns.
[0091] Furthermore, in this embodiment, the proxy model generation unit 125 generates a proxy model based on historical data of macro stress obtained as a result of performing multiple numerical material tests on the test piece (e.g., an example of the first data) and historical data saved as macro strain and macro accumulated strain in multiple numerical material tests (e.g., an example of the second data). This makes it possible to appropriately generate a proxy model based on the first data based on the macro stress obtained by performing multiple numerical material tests, and the second data based on the macro strain and macro accumulated strain that are stored as conditions for the multiple numerical material tests.
[0092] In this case, the surrogate model generation unit 125 may generate a surrogate model by machine learning using historical data (e.g., an example of the second data) stored as macro strain and macro accumulated strain in multiple numerical material tests as explanatory variable data, and historical data (e.g., an example of the first data) of macro stress obtained as a result of performing multiple numerical material tests on the above test piece as objective variable data. This makes it possible to appropriately generate a surrogate model by machine learning using the second data as explanatory variable data and the first data as response variable data.
[0093] Furthermore, in this embodiment, the surrogate model generating unit 125 optimizes model parameters including regularization parameters (such as the regularization parameter "η"). This makes it possible to appropriately generate a surrogate model while reducing the possibility of overfitting.
[0094] In this case, the surrogate model is a model based on the radial basis function (RBF), and the model parameters are parameters related to the radial basis function (parameter “β” and discrimination parameter “n ψ ", etc.) may be included. This allows optimizing model parameters, including parameters related to the radial basis functions, to generate a surrogate model based on the radial basis functions.
[0095] In this case, the parameters related to the radial basis functions are the discrimination parameters (discrimination parameters “n ψ ", etc.) (for example, an example of an identification parameter for identifying each of a plurality of predetermined radial basis functions). This makes it possible to generate a surrogate model based on radial basis functions by optimizing model parameters including discrimination parameters for discriminating between each of a plurality of predetermined radial basis functions. [Explanation of symbols]
[0096] 1. Information processing device 110 Macro structure acquisition unit 120 Homogenization analysis section 121 Microstructure Numerical Model Generation Unit 123 Numerical Material Testing Department 125 Surrogate Model Generation Unit 130 Macro Structure Analysis Department 140 Localization Analysis Unit
Claims
1. A generating device for generating a surrogate model of a homogenization constitutive law that enables a decoupled multi-scale analysis, comprising: A generating device including a generating unit that generates a representative model that represents a relationship between macro-stress and macro-accumulated strain based on the results of multiple numerical material tests performed on a test piece.
2. The generating unit generates the representative model based on a macro stress obtained by performing a numerical material test multiple times on the test piece, and a macro strain and a macro accumulated strain stored as conditions of the numerical material test. The generating device of claim 1 .
3. The generation unit has a calculation unit that calculates a macro cumulative strain based on a macro strain that is a condition of the numerical material test. The generating device of claim 1 .
4. The generation unit generates the representative model based on a result of a numerical material test performed under conditions of a plurality of different macro strain patterns. The generating device of claim 1 .
5. The generating unit generates the representative model based on first data based on macro stress obtained by performing a plurality of numerical material tests and second data based on macro strain and macro accumulated strain stored as conditions of the plurality of numerical material tests. The generating device of claim 1 .
6. The generation unit generates the surrogate model by machine learning using the second data as explanatory variable data and the first data as objective variable data. The generating device according to claim 5 .
7. The generation unit has an optimization unit that optimizes model parameters including a regularization parameter. The generating device according to claim 6.
8. the surrogate model is a radial basis function based model; The model parameters include parameters related to the radial basis functions. The generating device according to claim 7.
9. The parameters relating to the radial basis functions include identification parameters for identifying each of a plurality of predetermined radial basis functions. The generating device according to claim 8.
10. A method for generating a surrogate model of a homogenization constitutive law that enables a decoupled multi-scale analysis, comprising the steps of: A generation method including generating a representative model that represents the relationship between macro-stress and macro-accumulated strain based on the results of multiple numerical material tests performed on a test specimen.
11. A program for causing a computer to generate a surrogate model of a homogenization constitutive law that enables a decoupled multi-scale analysis, comprising: A program for causing the computer to function as a generating unit that generates a representative model that represents the relationship between macro-stress and macro-accumulated strain based on the results of multiple numerical material tests performed on a test piece.
Citation Information
Cited By
Intelligent analysis system and method for stress distribution of steel truss structure
CN119880226A
