Method and System for Predicting Macroscopic Physical Properties of Objects Based on Deep Learning

Abstract

The present disclosure provides a method and system for predicting the macroscopic physical properties of an object based on deep learning, including: obtaining the equivalent base material properties and voxelized microstructure units after deformation based on a given deformation mode and geometric microstructure; obtaining the displacement of each microstructure unit voxel under multi-directional unit strain based on a pre-trained deep learning model and the acquired equivalent base material properties and voxelized microstructure units after deformation; obtaining the physical property index of the deformed geometric microstructure based on the displacement; wherein the deep learning model is based on a U-Net network structure, including sequentially connected convolution blocks, upsampling blocks and convolution layers, and the convolution blocks include sequentially connected first convolution layers, first normalization processing modules, activation functions, second convolution layers, second normalization processing modules and activation functions; the input of the model is a four-dimensional vector, the first dimension is the equivalent base material properties after deformation, and the remaining three dimensions are voxelized microstructure units.

Description

Technical Field

[0001] The present disclosure belongs to the technical field of macroscopic physical property prediction of objects, and in particular, relates to a method and system for predicting macroscopic physical properties of objects based on deep learning. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Geometric microstructures are ubiquitous in natural objects, and almost all materials have microstructures of a certain scale to reflect their unique physical properties. By using different geometric microstructures, it can be applied to additive manufacturing fields such as machinery, aerospace, and civil engineering.

[0004] Since different geometric microstructures have their own unique geometric structures, volume fractions and other properties, for traditional methods, the calculation time is long and the calculation cost is high. In the past few decades, the homogenization method has gradually developed into a multi-scale technology, which relies on the morphology of the microstructure and solves the relationship between macro-scale stress and strain by solving micro-scale problems. The most popular homogenization calculation structure is a voxel-based dual-scale microstructure calculation structure, which first divides the object into a coarse-grained voxel grid and fills the grid with periodic geometric microstructures for simulation and optimization. However, the inventors found that although this dual-scale calculation framework has achieved a balance between design accuracy and simulation efficiency, it still faces the problem of huge simulation errors caused by the uneven edge voxels; a method to solve the problem of boundary voxel shape is to use hexahedrons instead of cubic voxels for micro-scale filling, but such a method requires that the shape parameters of the deformed microstructure be considered in the homogenization calculation, which greatly increases the calculation cost and increases the calculation error due to the introduction of shape parameters. Summary of the invention

[0005] In order to solve the above-mentioned problems, the present disclosure provides a method and system for predicting the macroscopic physical properties of objects based on deep learning. The scheme predicts the macroscopic physical properties of objects based on a constructed deep learning network model and a deformed geometric microstructure, and obtains macroscopic-scale physical properties in hundreds of times faster time than traditional numerical averaging methods, thereby achieving almost real-time prediction of material properties.

[0006] According to a first aspect of an embodiment of the present disclosure, a method for predicting macroscopic physical properties of an object based on deep learning is provided, comprising:

[0007] Based on the given deformation mode and geometric microstructure, the equivalent base material properties and voxelized microstructure units after deformation are obtained;

[0008] Based on the pre-trained deep learning model and the acquired deformed equivalent base material properties and voxelized microstructure units, the displacement of each microstructure unit voxel under multi-directional unit strain is obtained;

[0009] Based on the obtained displacement, the corresponding homogenized constitutive matrix is ​​calculated to obtain the physical property index of the deformed geometric microstructure;

[0010] Among them, the deep learning model is based on the U-Net network structure, including sequentially connected convolution blocks, upsampling blocks and convolution layers, and the convolution blocks include sequentially connected first convolution layers, first normalization processing modules, activation functions, second convolution layers, second normalization processing modules and activation functions; the input of the model is a four-dimensional vector, the first dimension is the equivalent base material properties after deformation, and the remaining three dimensions are voxelized microstructure units.

[0011] Furthermore, the training of the deep learning model is specifically as follows:

[0012] Constructing a training data set, wherein the samples in the data set include a voxelized expression matrix of each deformed microstructure, a corresponding deformation parameter expression matrix, and an elastic tensor matrix describing the properties of a base material;

[0013] The constructed deep learning model is trained based on the training data set to obtain a trained deep learning model.

[0014] Furthermore, the deformed geometric microstructure is represented by a binary group consisting of a cubic microstructure and a deformation matrix. The cubic microstructure determines an optimal parallelepiped matching the deformed hexahedron based on an iterative closest point algorithm.

[0015] Furthermore, an isotropic material is selected as a base material, and the base material properties are obtained based on the Young's modulus and Poisson's ratio of the material. An equivalent transformation is performed based on the obtained base material properties to obtain equivalent base material properties after deformation.

[0016] Furthermore, the equivalent transformation is specifically expressed as follows:

[0017]

[0018] Among them, C H is the elastic tensor of the microstructure before transformation, F represents the mapping relationship from cubic units to hexahedral units, is the equivalent elastic tensor of the deformed microstructure after transformation.

[0019] Furthermore, the loss function used in the deep learning model training process is constructed based on the minimum potential energy principle, which is specifically expressed as follows:

[0020]

[0021] Where u is the predicted displacement, u T is the transpose of the displacement matrix, K is the stiffness matrix, f is the load, where subscript i=1,2...6.

[0022] Furthermore, the physical property indicators include but are not limited to stress-strain relationship, stress-strain distribution, yield strength and shear strength.

[0023] According to a second aspect of an embodiment of the present disclosure, a system for predicting macroscopic physical properties of an object based on deep learning is provided, comprising:

[0024] A data acquisition unit, which is used to obtain the equivalent base material properties and voxelized microstructure units after deformation based on a given deformation mode and geometric microstructure;

[0025] A strain displacement acquisition unit, which is used to obtain the displacement of each microstructure unit voxel under multi-directional unit strain based on a pre-trained deep learning model and the acquired deformed equivalent base material properties and voxelized microstructure units;

[0026] A physical property acquisition unit, which is used to calculate the corresponding homogenized constitutive matrix based on the obtained displacement, and obtain the physical property index of the deformed geometric microstructure;

[0027] Among them, the deep learning model is based on the U-Net network structure, including sequentially connected convolution blocks, upsampling blocks and convolution layers, and the convolution blocks include sequentially connected first convolution layers, first normalization processing modules, activation functions, second convolution layers, second normalization processing modules and activation functions; the input of the model is a four-dimensional vector, the first dimension is the equivalent base material properties after deformation, and the remaining three dimensions are voxelized microstructure units.

[0028] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which a program is stored, and when the program is executed by a processor, a method for predicting macroscopic physical properties of an object based on deep learning as described above is implemented.

[0029] According to a fourth aspect of an embodiment of the present invention, there is provided an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, a method for predicting macroscopic physical properties of an object based on deep learning as described above is implemented.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] (1) The present disclosure provides a method and system for predicting the macroscopic physical properties of an object based on deep learning. The scheme predicts the macroscopic physical properties of an object based on a constructed deep learning network model and a deformed geometric microstructure. Compared with the traditional numerical averaging method, the macroscopic-scale physical properties are obtained in hundreds of times faster time, thus achieving almost real-time prediction of material properties.

[0032] (2) The scheme provides a machine learning training method that does not require dataset labels, saving the huge computational cost of obtaining dataset labels;

[0033] (3) The scheme provides a more generalized network training idea, which can calculate more mechanical performance indicators, such as stress and strain distribution, yield strength, shear strength, etc., in addition to predicting the macro-scale elastic tensor matrix.

[0034] Advantages of additional aspects of the present disclosure will be given in part in the following description and in part will become apparent from the following description or will be learned through practice of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0036] Figure 1 This is a flow chart of the method for predicting macroscopic physical properties of an object based on deep learning described in the first embodiment of the present disclosure;

[0037] Figure 2 This is a schematic diagram of the deep learning model architecture described in Embodiment 1 of the present disclosure;

[0038] Figure 3(a) to Figure 3(b) This is a schematic diagram of the shape-material transformation described in the first embodiment of the present disclosure;

[0039] Figure 4(a) to Figure 4(c) They are respectively display diagrams of the prediction results described in the first embodiment of the present disclosure. DETAILED DESCRIPTION

[0040] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0041] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0042] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0043] In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other.

[0044] Embodiment 1:

[0045] The purpose of this embodiment is to provide a method for predicting the macroscopic physical properties of an object based on deep learning.

[0046] like Figure 1 As shown, a method for predicting macroscopic physical properties of an object based on deep learning includes:

[0047] Based on the given deformation mode and geometric microstructure, the equivalent base material properties and voxelized microstructure units after deformation are obtained;

[0048] Based on the pre-trained deep learning model and the acquired deformed equivalent base material properties and voxelized microstructure units, the displacement of each microstructure unit voxel under multi-directional unit strain is obtained;

[0049] Based on the obtained displacement, the corresponding homogenized constitutive matrix is ​​calculated to obtain the physical property index of the deformed geometric microstructure;

[0050] Among them, the deep learning model is based on the U-Net network structure, including sequentially connected convolution blocks, upsampling blocks and convolution layers, and the convolution blocks include sequentially connected first convolution layers, first normalization processing modules, activation functions, second convolution layers, second normalization processing modules and activation functions; the input of the model is a four-dimensional vector, the first dimension is the equivalent base material properties after deformation, and the remaining three dimensions are voxelized microstructure units.

[0051] Furthermore, the training of the deep learning model is specifically as follows:

[0052] Construct a training data set, which includes 62,500 deformed microstructure data samples. Each sample data is a voxel expression matrix of each deformed microstructure and a corresponding deformation parameter expression matrix and an elastic tensor matrix describing the properties of the base material. The size of each matrix is ​​36×36×36, 3×3, 6×6, and is stored in a data format with the suffix name mat. The deformation matrix F = ST, the shear matrix S (α xy ,α yz,α zx ) is expressed as the shearing along the x-axis to the y-direction, the shearing along the y-axis to the z-direction, and the shearing along the z-axis to the x-direction; the scaling matrix T(l x ,l y ,l z ) represents the scaling along the x, y, and z axes respectively;

[0053] The constructed deep learning model is trained based on the training data set to obtain a trained deep learning model.

[0054] Furthermore, the deformed geometric microstructure is represented by a binary group consisting of a cubic microstructure and a deformation matrix. The cubic microstructure determines an optimal parallelepiped matching the deformed hexahedron based on an iterative closest point algorithm.

[0055] Furthermore, an isotropic material is selected as a base material, and the base material properties are obtained based on the Young's modulus and Poisson's ratio of the material. An equivalent transformation is performed based on the obtained base material properties to obtain equivalent base material properties after deformation.

[0056] Furthermore, the physical property indicators include but are not limited to stress-strain relationship, stress-strain distribution, yield strength and shear strength.

[0057] Specifically, for ease of understanding, the solution disclosed in the present disclosure is described in detail below with reference to the accompanying drawings:

[0058] The present disclosure provides a method for predicting the macroscopic physical properties of an object based on deep learning. For a specific deformation mode and geometric microstructure given by a user, the deep learning network of deep learning is used to calculate the displacement of each microscopic grid voxel under unit strain in six directions, and then obtains the stress-strain relationship, stress-strain distribution, yield strength, shear strength and other physical property indicators of the deformed geometric microstructure. The specific process includes: deep learning model construction, loss function construction, training data set construction, model training and physical property prediction. Each process is described in detail below:

[0059] (1) Deep learning model construction

[0060] like Figure 2 As shown in FIG, the deep learning model is improved based on the existing U-Net network. Figure 2 As shown, the network model input size is 36 4 , where the first dimension is the equivalent base material property after deformation, the last three dimensions are voxelized microstructure units with a resolution of 36, and the network output is 18×36 3, represents the displacement of each voxel under unit strain in six directions. The voxelized microstructure is filled with the Tubular Gyroid (TG) type in the three-periodic minimal surface to generate a voxelized microstructure, and 40 random samples are taken from the range of volume fraction 2.4%-33% as the basic microstructure unit. The deformation matrix is ​​applied to the basic microstructure unit to obtain 62,500 deformed microstructure units as training data.

[0061] Furthermore, the deep learning model mainly includes three convolution blocks, two groups of upsampling blocks and a convolution layer at the end.

[0062] The number of input channels of the three convolution blocks are 36, 64, and 128 respectively, and the number of output channels are 64, 128, and 256 respectively. There is a maximum pooling layer with a kernel size of 2 and a stride of 2 between every two convolution blocks.

[0063] Each convolution block consists of six modules, namely, the convolution layer (i.e., the first convolution layer) (the convolution kernel size is 3, the step size is 1, the number of padding edges is 1, and the bias mode is turned on), the BatchNorm normalization processing module (i.e., the first normalization processing module), the ReLU activation function (overlay is turned on), the convolution layer (i.e., the second convolution layer) (the number of input channels is the same as the number of output channels), the BatchNorm normalization processing module (i.e., the second normalization processing module) (same as above), and the ReLU activation function (same as above).

[0064] A set of upsampling blocks consists of an upsampling layer and a convolutional layer (e.g. Figure 1 The input of the convolution layer comes from the result of merging the upsampling layer and the convolution block result from the arrow in the channel dimension). The upsampling layer includes an upsampling sublayer (input-output ratio is 2), a convolution layer (convolution kernel size is 3, step size is 1, interval is 1, bias mode is turned on), a BatchNorm module and a ReLU activation function (same as above). The input channels of the two groups of convolution layers are 256 and 128 respectively, and the output channels are 128 and 64 respectively.

[0065] After the last convolutional layer obtains 64-channel input, it obtains 18-channel results, and the other parameters are the same as above.

[0066] (2) Loss function construction

[0067] Different from the commonly used labeled training method, the scheme disclosed in the present invention constructs a new deep learning loss function based on the minimum potential energy principle, so that the training process no longer relies on labels and becomes unsupervised learning. The minimum potential energy principle is one of the energy principles of elasticity mechanics, and can serve as an important basis for direct solutions and finite element calculations in elasticity mechanics. In summary, the principle of minimum potential energy is: among all possible deformation displacement fields, the real displacement field makes the total potential energy functional take the minimum value. Substituting the displacement output by the network into the loss function to participate in the calculation and back-propagation process, the network can eventually converge to predict the real displacement field, and perform subsequent calculations through the displacement field. The unsupervised training process can be achieved using a loss function based on the minimum potential energy principle.

[0068] We set the loss function to

[0069]

[0070] Where u is the predicted displacement, u T is the transpose of the displacement matrix, K is the stiffness matrix, f is the load, where the subscript i = 1, 2...6, and its gradient It can be expressed as This method reduces the energy difference between the two energies, making the displacement u predicted by the network more accurate and saving the time of calculating the dataset labels.

[0071] (3) Construction of training dataset

[0072] In order to better place the deformed geometric microstructure into the deep learning network (otherwise the convolutional layer, pooling layer, etc. of the convolutional neural network cannot work properly on the deformed input), we generate a <cube microstructure, deformation matrix> tuple to store the deformed geometric microstructure instead. We store the undeformed cube microstructure as a voxel tensor of a specific resolution, and for the deformed data we use the iterative closest point algorithm to find the optimal parallelepiped match from the deformed hexahedron. We simulate a solved parallelepiped by determining the elongation and compression ratio of the cube on the x, y, and z axes and the shear angles of the x-axis to the y-axis, the y-axis to the z-axis, and the x-axis to the z-axis.

[0073] In constructing the dataset, we used the isotropic material C b (E,v) (The superscript b indicates that this elastic tensor is the elastic tensor describing the base material, which is used to distinguish the C obtained by homogenization calculation H, E represents Young's modulus, v represents Poisson's ratio) as the base material, the material properties can be easily calculated through Young's modulus E and Poisson's ratio v. Since only solid nodes are used as input into the deep learning network in machine learning, some weights in the network lack backward gradients, the network parameters are not updated, and the network output is wrong. When constructing the data set, we set the geometric microstructure to be a composite of hard and soft materials, where the properties of the hard material are set to (E=1,v=0.3), and the properties of the soft material are set to (E=10 -6 ,v=0.3), the transformation formula provided by the material migration method of the base material Among them, using C b (E b ,v b )and Represents two different base material properties. The property migration of the base material can be expressed as H (E i ,v i,j ,G i,j ) to i, j∈{x,y,z} represents the direction of the coordinate axis, G=F -1 Representing the inverse of the deformation matrix, we can transform the network output results to obtain the macroscopic material physical properties of the actual input material properties.

[0074] When the <cube microstructure, deformation matrix> tuple is used instead of the deformed geometric microstructure as input to the deep learning network for training, the deformation that should be applied to the cube microstructure needs to be converted into the deformation of the base material properties. Calculate, where G = F -1 is the inverse mapping from parallelepiped to cube. The elastic tensor of the deformed base material The cube unit microstructure is used as the input of the neural network for network training and prediction.

[0075] The solution is to set the x-axis, y-axis, and z-axis telescopic ranges and the angle range between the two axes [75°, 90°], the training microstructure model Tubular Gyroid and its volume fraction range [2.4%, 33%], and 10% of the samples were randomly selected from the sample space (the space size is 62500) for training.

[0076] (4) Model training

[0077] We use the MATLAB version of the homogenization code as the ground truth value of the network generated results The result C calculated by the deep learning network output H , through the formula Calculate the relative error to determine how close the prediction result is to the true value.

[0078] During the training process, we randomly shuffled the samples and divided the training set and the test set into 8:2. The Adam optimizer was used as the training optimizer, and the training learning rate was set to 10. -4 , the batch size is set to 8.

[0079] (5) Physical property prediction

[0080] 1) The main process of using the trained deep learning model is:

[0081] 2) Through the formula Compute the elastic tensor for hard materials The elastic tensor corresponding to the soft material by Construct the network input, where Ω is the microstructure represented by a voxelized matrix and I is the identity matrix.

[0082] 3) Obtaining microscopic displacements by using a trained deep learning model Compute the homogenized elastic tensor matrix

[0083] Specifically, the specific steps are as follows:

[0084] The displacement field output by network prediction can be used to calculate the corresponding homogenized constitutive matrix according to the following formula:

[0085]

[0086] in, is the calculated homogenized constitutive matrix, Ω represents the voxelized microstructure, |Ω| is the volume of the corresponding microstructure, and u 0 =K e / f e represents the displacement of each voxel unit, μ represents the displacement field output by the network prediction, and subscript e=1...N 3 Represents a certain voxel unit;

[0087] By formula Restored to the elastic tensor matrix C of the deformed geometric microstructure H ;

[0088] The physical property indicators including Young's modulus, shear modulus, and Poisson's ratio can be directly obtained from the inverse matrix of the constitutive matrix. The stress-strain distribution, yield strength and other indicators can be obtained according to the existing calculation formula. For details, please refer to the method described in the paper Liu, Peiqing, et al. "Mechanical property profiles of microstructures via asymptotic homogenization." Computers & Graphics for calculation.

[0089] Further, such as Figure 3(a) to Figure 3(b) The figure shows the shape-material transformation schematic diagram described in the first embodiment of the present disclosure, wherein FIG3(a) shows that in the data set construction, the deformation of the deformed microstructure is transformed into the deformation of the base material properties through the shape-material transformation step, thereby ensuring that the network input data is the voxel data of the standard cubic unit and the properties of the corresponding deformed base material. FIG3(b) shows that after the properties of the cubic unit microstructure are calculated by the displacement predicted by the network, the deformation is transformed from the tensor matrix to the deformed microstructure through the inverse process of the shape-material transformation, thereby completing the property prediction of the deformed microstructure.

[0090] Further, such as Figure 4(a) to Figure 4(c) Shown is a display diagram of the prediction results described in Example 1 of the present disclosure, in which the three deformed TPMS-TG microstructures have the same volume fraction (10%), but different shape parameters (left column). From top to bottom, their shape parameters are (1, 1, 1, 75°, 75°, 75°), (2 / 3, 2 / 3, 3 / 2, 90°, 90°, 90°), (2 / 3, 2 / 3, 3 / 2, 75°, 75°, 75°), respectively representing the scaling along the x / y / z axis and the misalignment along the x-axis to the y-direction, the misalignment along the y-axis to the z-direction, and the misalignment along the z-axis to the x-direction. The second and third columns are graphs of the predicted Young's modulus given by numerical homogenization (middle) and DH-Net (right). It can be found that the results calculated by the numerical homogenization method are very close to the results predicted using DH-Net, which proves the effectiveness of DH-Net.

[0091] Embodiment 2:

[0092] The purpose of this embodiment is to provide a system for predicting macroscopic physical properties of objects based on deep learning, including:

[0093] A data acquisition unit, which is used to obtain the equivalent base material properties and voxelized microstructure units after deformation based on a given deformation mode and geometric microstructure;

[0094] A strain displacement acquisition unit, which is used to obtain the displacement of each microstructure unit voxel under multi-directional unit strain based on a pre-trained deep learning model and the acquired deformed equivalent base material properties and voxelized microstructure units;

[0095] A physical property acquisition unit, which is used to calculate the corresponding homogenized constitutive matrix based on the obtained displacement, and obtain the physical property index of the deformed geometric microstructure;

[0096] Among them, the deep learning model is based on the U-Net network structure, including sequentially connected convolution blocks, upsampling blocks and convolution layers, and the convolution blocks include sequentially connected first convolution layers, first normalization processing modules, activation functions, second convolution layers, second normalization processing modules and activation functions; the input of the model is a four-dimensional vector, the first dimension is the equivalent base material properties after deformation, and the remaining three dimensions are voxelized microstructure units.

[0097] In further embodiments, there is also provided:

[0098] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, no further description is given here.

[0099] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0100] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0101] A computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the method described in embodiment 1 is completed.

[0102] The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0103] Those skilled in the art will appreciate that the units, i.e., algorithm steps, of the various examples described in conjunction with this embodiment can be implemented in electronic hardware or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0104] The method and system for predicting the macroscopic physical properties of an object based on deep learning provided in the above-mentioned embodiments can be implemented and have broad application prospects.

[0105] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for predicting the macroscopic physical properties of an object based on deep learning, characterized in that, it includes: Based on the given deformation method and geometric microstructure, obtain the equivalent base material properties after deformation and the voxelized microstructure units; Among them, the deformation matrix , the shear matrix represents the shear in the y direction along the x axis, the shear in the z direction along the y axis, and the shear in the x direction along the z axis; the scaling matrix represents the scaling along the x, y, and z axes respectively; Using an isotropic material As the base material, a dataset is constructed by calculating the material properties from the Young's modulus E and the Poisson's ratio v. Here, the superscript indicates that this elastic tensor is the elastic tensor describing the base material, used to distinguish it from the one obtained by homogenization calculation ; When constructing the dataset, it is set that the geometric microstructure is formed by the composite of hard material and soft material. Among them, the properties of the hard material are set to , and the properties of the soft material are set to . Through the transformation formula provided by the material migration method of the base material , where and represent the properties of two different base materials, and the property migration of the base material is expressed as from transformed to . represents the coordinate axis direction, represents the inverse of the deformation matrix. By transforming the network output result, the macroscopic material physical properties of the actual input material properties are obtained; Based on the pre-trained deep learning model and the obtained equivalent base material properties and voxelized microstructure units after deformation, obtain the displacements of each voxel of the microstructure unit under multi-directional unit strains; Based on the obtained displacements, calculate the corresponding homogenized constitutive matrix to obtain the physical property indexes of the deformed geometric microstructure; Among them, the deep learning model is based on the U-Net network structure, including a sequentially connected convolutional block, an upsampling block, and a convolutional layer. The convolutional block includes a sequentially connected first convolutional layer, a first normalization processing module, an activation function, a second convolutional layer, a second normalization processing module, and an activation function; the input of the model is a four-dimensional vector, the first dimension is the equivalent base material properties after deformation, and the remaining three dimensions are the voxelized microstructure units.

2. A method for predicting the macroscopic physical properties of an object based on deep learning according to claim 1, characterized in that, the training of the deep learning model is specifically: Construct a training data set, and the samples in the data set include the voxelized expression matrix of each deformed microstructure, the corresponding deformed parameter expression matrix, and the elastic tensor matrix describing the properties of the base material; Based on the training data set, train the constructed deep learning model to obtain a trained deep learning model.

3. A method for predicting the macroscopic physical properties of an object based on deep learning according to claim 1, characterized in that, For the deformed geometric microstructure, it is represented by a binary group composed of a cubic microstructure and a deformation matrix. The cubic microstructure is based on the iterative closest point algorithm to determine the optimal parallelepiped matching it from the deformed hexahedron.

4. A method for predicting the macroscopic physical properties of an object based on deep learning according to claim 1, characterized in that, Select an isotropic material as the base material, obtain the base material properties based on the Young's modulus and Poisson's ratio of the material, and perform equivalent transformation based on the obtained base material properties to obtain the equivalent base material properties after deformation.

5. A method for predicting the macroscopic physical properties of an object based on deep learning according to claim 4, characterized in that, the equivalent transformation is specifically expressed as follows: Among them, is the elastic tensor of the microstructure before transformation, represents the mapping relationship from the cubic element to the hexahedral element, is the equivalent elastic tensor of the deformed microstructure after transformation.

6. A method for predicting the macroscopic physical properties of an object based on deep learning according to claim 1, characterized in that, In the training process of the deep learning model, the loss function is constructed based on the principle of minimum potential energy, and is specifically expressed as follows: wherein, u is the predicted displacement, u T is the transpose of the displacement matrix, is the stiffness matrix, is the load, where the subscript .

7. A method for predicting the macroscopic physical properties of an object based on deep learning according to claim 1, characterized in that, The physical property indexes include but are not limited to the stress-strain relationship, stress-strain distribution, yield strength, and shear strength.

8. A system for predicting the macroscopic physical properties of an object based on deep learning, characterized in that, it includes: A data acquisition unit, which is used to obtain the equivalent base material properties after deformation and the voxelized microstructure units based on the given deformation method and geometric microstructure; Among them, the deformation matrix , the shear matrix represents the shear in the y direction along the x axis, the shear in the z direction along the y axis, and the shear in the x direction along the z axis; the scaling matrix represents the scaling along the x, y, and z axes respectively; Using isotropic materials As the base material, the material properties are calculated from the Young's modulus E and Poisson's ratio v to construct a data set. Here, the superscript indicates that this elastic tensor describes the elastic tensor of the base material, used to distinguish it from the one obtained by homogenization calculation ; When constructing the data set, it is set that the geometric microstructure is formed by the composite of hard material and soft material. Among them, the properties of the hard material are set to , and the properties of the soft material are set to . Through the transformation formula provided by the material migration method of the base material , where and represent the properties of two different base materials. The property migration of the base material is expressed as the transformation from to . represents the coordinate axis direction, represents the inverse of the deformation matrix. By transforming the network output result, the macroscopic material physical properties of the actual input material properties are obtained; A strain displacement acquisition unit, which is configured to obtain the displacements of each voxel of the microstructure unit under multi-directional unit strains based on a pre-trained deep learning model and the obtained deformed equivalent base material properties and voxelized microstructure units; A physical property acquisition unit, which is configured to calculate the corresponding homogenized constitutive matrix based on the obtained displacements and obtain the physical property indexes of the deformed geometric microstructure; Wherein, the deep learning model is based on a U-Net network structure and includes a sequentially connected convolutional block, an upsampling block, and a convolutional layer. The convolutional block includes a sequentially connected first convolutional layer, a first normalization processing module, an activation function, a second convolutional layer, a second normalization processing module, and an activation function; the input of the model is a four-dimensional vector, the first dimension is the deformed equivalent base material property, and the remaining three dimensions are voxelized microstructure units.

9. A computer-readable storage medium, on which a program is stored, Characterized in that, When the program is executed by a processor, it implements a method for predicting the macroscopic physical properties of an object based on deep learning as described in any one of claims 1-7.

10. An electronic device, comprising a memory, a processor, and a program stored on the memory and executable on the processor, Characterized in that, When the processor executes the program, it implements a method for predicting the macroscopic physical properties of an object based on deep learning as described in any one of claims 1-7.

Images