Ultrahigh-speed collision process simulation method and device based on neural network

Abstract

The embodiment of the invention provides an ultra-high-speed collision process simulation method and device based on a neural network, and the method comprises the steps: obtaining the space coordinate information of a target substance point at a target moment, and enabling the target substance point to be a substance point of a to-be-simulated object in the ultra-high-speed collision process; inputting the space coordinate information into a pre-trained target neural network to obtain initial equivalent plastic strain and initial stress triaxial degree output by the target neural network; inputting the space coordinate information and the initial equivalent plastic strain into a pre-trained first operator neural network to obtain the stress triaxial degree of the target substance point output by the first operator neural network at the target moment; and inputting the space coordinate information and the initial stress three-axis degree into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target substance point at the target moment, wherein the equivalent plastic strain is output by the second operator neural network. According to the method, high-precision simulation and prediction can be realized in an ultra-high-speed collision complex environment.

Description

Technical Field

[0001] The present disclosure relates to the field of deep learning technology, and in particular to a method and device for simulating a hypervelocity collision process based on a neural network. Background Art

[0002] During a hypervelocity collision, materials usually undergo drastic deformation, which causes traditional grid algorithms to face grid distortion problems, which in turn leads to a decrease in calculation accuracy and a reduction in time step, affecting the efficiency of the calculation. Particle-based methods or particle-based grid methods combined with particle-based methods can effectively overcome the grid distortion problem and are suitable for processing materials with large deformations. However, these methods have high computational costs and still face great challenges in numerically simulating complex physical problems due to limited data. Therefore, there is an urgent need for a method to achieve high-precision simulation and prediction in complex environments such as hypervelocity collisions. Summary of the invention

[0003] In view of this, the embodiments of the present disclosure provide at least one method and device for simulating a hypervelocity collision process based on a neural network, so as to improve the simulation accuracy of a hypervelocity collision process such as a Taylor rod impact.

[0004] Specifically, the embodiments of the present disclosure are implemented through the following technical solutions:

[0005] In a first aspect, a method for simulating a hypervelocity collision process based on a neural network is provided, the method comprising:

[0006] Acquire spatial coordinate information of a target material point at a target time, wherein the target material point is a material point of an object to be simulated during a hypervelocity collision;

[0007] Inputting the spatial coordinate information into a pre-trained target neural network to obtain the initial equivalent plastic strain and initial stress triaxiality output by the target neural network;

[0008] Inputting the spatial coordinate information and the initial equivalent plastic strain into a pre-trained first operator neural network to obtain the stress triaxiality of the target material point at the target time output by the first operator neural network;

[0009] Inputting the spatial coordinate information and the initial stress triaxiality into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target material point at the target time output by the second operator neural network;

[0010] The combination of the target neural network, the first operator neural network and the second operator neural network is used to simulate the nonlinear mapping relationship between the spatial coordinate information of material points, stress triaxiality and equivalent plastic strain during a hypervelocity collision.

[0011] In combination with any implementation manner provided in the present disclosure, the first operator neural network and the second operator neural network adopt the same network structure.

[0012] In combination with any embodiment provided in the present disclosure, the first operator neural network adopts a DeepONet network structure; the first operator neural network is composed of a first branch network and a first trunk network;

[0013] The step of inputting the spatial coordinate information and the initial equivalent plastic strain into a pre-trained first operator neural network to obtain the stress triaxiality of the target material point at the target time output by the first operator neural network includes:

[0014] Inputting the initial equivalent plastic strain into the first branch network to extract an initial equivalent plastic strain eigenvector;

[0015] Inputting the spatial coordinate information into a first backbone network to extract a first spatial coordinate feature vector;

[0016] The simulated stress triaxiality of the target material point at the target time is obtained according to the initial equivalent plastic strain eigenvector and the first space coordinate eigenvector.

[0017] In combination with any embodiment provided in the present disclosure, the second operator neural network adopts a DeepONet network structure; the second operator neural network is composed of a second branch network and a second trunk network;

[0018] The step of inputting the spatial coordinate information and the initial stress triaxiality into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target material point at the target time output by the second operator neural network includes:

[0019] Inputting the initial stress triaxiality into the second branch network to extract the initial stress triaxiality feature vector;

[0020] Inputting the spatial coordinate information into a second backbone network to extract a second spatial coordinate feature vector;

[0021] According to the initial stress triaxiality eigenvector and the second spatial coordinate eigenvector, the simulated equivalent plastic strain of the target material point at the target time is obtained.

[0022] In combination with any embodiment provided in the present disclosure, the first operator neural network is trained in the following manner:

[0023] Inputting the sample space coordinate information and the equivalent plastic strain measured value of the sample material point into the initial first operator neural network to be trained, and obtaining the stress triaxiality simulation value output by the initial first operator neural network;

[0024] determining a first network loss according to a difference between the stress triaxiality simulation value and the stress triaxiality measured value;

[0025] The network parameters of the initial first operator neural network are adjusted according to the first network loss, and when the training end condition is reached, the first operator neural network is obtained.

[0026] In combination with any embodiment provided in the present disclosure, the second operator neural network is trained in the following manner:

[0027] Inputting the sample space coordinate information of the sample material point and the measured value of stress triaxiality into the initial second operator neural network to be trained, and obtaining the equivalent plastic strain simulation value output by the initial second operator neural network;

[0028] Determining a second network loss according to a difference between the equivalent plastic strain simulation value and the equivalent plastic strain measured value;

[0029] The network parameters of the initial second operator neural network are adjusted according to the second network loss, and when the training end condition is reached, the second operator neural network is obtained.

[0030] In combination with any of the embodiments provided in the present disclosure, the target neural network is trained in the following manner:

[0031] Inputting sample space coordinate information of sample material points into an initial neural network to obtain an initial value of equivalent plastic strain and an initial value of stress triaxiality output by the initial neural network;

[0032] Inputting the sample space coordinate information and the initial value of the equivalent plastic strain into a pre-trained first operator neural network to obtain a stress triaxiality prediction value of the sample material point output by the first operator neural network;

[0033] Inputting the sample space coordinate information and the initial value of the stress triaxiality into a pre-trained second operator neural network to obtain an equivalent plastic strain prediction value of the sample material point output by the second operator neural network;

[0034] Determine a third network loss according to a difference between the measured value of the stress triaxiality and the initial value of the stress triaxiality, and a difference between the measured value of the equivalent plastic strain and the initial value of the equivalent plastic strain;

[0035] Determine a fourth network loss according to a difference between the measured value of the stress triaxiality and the predicted value of the stress triaxiality, and a difference between the measured value of the equivalent plastic strain and the predicted value of the equivalent plastic strain;

[0036] The network parameters of the initial neural network are adjusted according to the third network loss and the fourth network loss, and when the training end condition is reached, the target neural network is obtained.

[0037] In a second aspect, a hypervelocity collision process simulation device based on a neural network is provided, the device comprising:

[0038] A data acquisition module, used to acquire spatial coordinate information of a target material point at a target time, wherein the target material point is a material point of an object to be simulated during a hypervelocity collision;

[0039] A target network module, used for inputting the spatial coordinate information into a pre-trained target neural network to obtain the initial equivalent plastic strain and initial stress triaxiality output by the target neural network;

[0040] A first operator module, used for inputting the spatial coordinate information and the initial equivalent plastic strain into a pre-trained first operator neural network, and obtaining the stress triaxiality of the target material point at the target time output by the first operator neural network;

[0041] A second operator module is used to input the spatial coordinate information and the initial stress triaxiality into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target material point at the target time output by the second operator neural network;

[0042] The combination of the target neural network, the first operator neural network and the second operator neural network is used to simulate the nonlinear mapping relationship between the spatial coordinate information of material points, stress triaxiality and equivalent plastic strain during a hypervelocity collision.

[0043] In a third aspect, an electronic device is provided, comprising a memory and a processor, wherein the memory is used to store computer instructions executable on the processor, and the processor is used to implement the neural network-based hypervelocity collision process simulation method described in any embodiment of the present disclosure when executing the computer instructions.

[0044] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the method for simulating a hypervelocity collision process based on a neural network as described in any embodiment of the present disclosure is implemented.

[0045] The technical solution of the disclosed embodiment provides a method for simulating ultra-high-speed collision processes based on a neural network. The neural network model obtained by pre-training through an operator learning method based on deep learning adopts a parallel two-way prediction mechanism, which can enhance the adaptability of the model in complex scenarios. It can learn the changing trends of various variables when the material is deformed during the collision process under limited measurement data. Therefore, when simulating the ultra-high-speed collision process, the stress triaxiality and equivalent plastic strain of the target material point at the target moment can be accurately and quickly simulated based on the spatial coordinates of the target material point at the target moment, overcoming the limitations of traditional numerical methods in large deformation and high nonlinear problems, effectively improving the simulation accuracy of ultra-high-speed collision processes such as Taylor bar impact, providing a more efficient solution for engineering simulation, and helping to promote research and application development in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in one or more embodiments of the present disclosure or related technologies, the drawings required for use in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in one or more embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0047] Figure 1 is a schematic diagram of a network structure of a neural network integrating physical information shown in at least one embodiment of the present disclosure;

[0048] Figure 2 is a schematic diagram of the architecture of DeepONet shown in at least one embodiment of the present disclosure;

[0049] Figure 3 is a schematic diagram of a training process of a target neural network shown in at least one embodiment of the present disclosure;

[0050] Figure 4 is a flow chart of a method for simulating a hypervelocity collision process based on a neural network according to at least one embodiment of the present disclosure;

[0051] Figure 5 is a flow chart of establishing a Taylor rod collision process simulation model according to at least one embodiment of the present disclosure;

[0052] Figure 6 is a simulation result diagram of a Taylor rod collision process shown in at least one embodiment of the present disclosure;

[0053] Figure 7 is a block diagram of a hypervelocity collision process simulation device based on a neural network according to at least one embodiment of the present disclosure;

[0054] Figure 8 It is a schematic diagram of the hardware structure of an electronic device shown in at least one embodiment of the present disclosure. DETAILED DESCRIPTION

[0055] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this specification. Instead, they are merely examples of devices and methods consistent with some aspects of this specification as detailed in the appended claims.

[0056] The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit this specification. The singular forms "a", "the" and "the" used in this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0057] It should be understood that although the terms first, second, third, etc. may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this specification, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0058] Hypervelocity collisions involve complex physical phenomena, such as shock waves, fluid dynamics and material behavior under ultra-high temperatures and high pressures, and usually involve complex physical problems, such as large deformation of materials and high nonlinearity, which require high-precision numerical simulations. The inventors have found that deep learning performs well in solving PDEs (Partial Differential Equations) and discovering physical laws, and can effectively handle nonlinear and high-dimensional PDE solving problems. The physics-driven PINN (Physics-informed Neural Network) method integrates physical information into training to reduce label dependence, while operator learning (such as DeepONet) is applied to solving complex fluid physics problems by learning nonlinear operators, bringing new ideas to engineering simulation.

[0059] Based on this, the embodiment of the present disclosure proposes a method for simulating a hypervelocity collision process based on a neural network. Through an operator learning method based on deep learning, a neural network model with a parallel bidirectional prediction mechanism is constructed. With limited measurement data, the changing trends of various variables when the material is deformed during the collision process can be learned. Therefore, when simulating the hypervelocity collision process, the stress triaxiality and equivalent plastic strain of the target material point can be accurately simulated only based on the spatial coordinates of the target material point, thereby effectively improving the simulation accuracy and speed of the hypervelocity collision process.

[0060] First, the technical terms involved in the embodiments of the present disclosure are explained.

[0061] Material point: The material point of the present disclosure adopts the definition of material point in the material point method. In the material point method, various information of the object is stored in discrete material points, that is, the object is discretized into material points. For example, during the collision of the Taylor rod, the rod itself can be regarded as a continuous object composed of several material points. During the hypervelocity collision, the object will produce a large deformation. As the object changes, the position of the material point will also change accordingly. The spatial coordinate information can be the three-dimensional spatial coordinates of the material point.

[0062] Equivalent plastic strain: refers to the value of plastic strain that is the same as that produced by uniaxial tension under a complex three-dimensional stress state, and is usually used to quantify the degree of plastic deformation of an object.

[0063] Stress triaxiality: It is defined as the hydrostatic pressure ratio Mises equivalent stress and is a quantity that measures the stress state of an object when it is subjected to force.

[0064] The network structure of the neural network model constructed in the embodiment of the present disclosure is explained below. For the convenience of description, this embodiment refers to the neural network model as a fusion physical information neural network.

[0065] Figure 1 An example of a fusion physical information neural network 10 is provided. The fusion physical information neural network 10 may include: a target neural network 11, a first operator neural network 12 and a second operator neural network 13.

[0066] The target neural network 11 is used to input the spatial coordinate information of the target material point and output the initial equivalent plastic strain and initial stress triaxiality of the target material point. This embodiment does not limit the network structure used by the target neural network 11. For example, DNN (Deep Neural Network), NN (Neural Network) and other network structures can be used.

[0067] The first operator neural network 12 is used to input the spatial coordinate information of the target material point and the initial equivalent plastic strain output by the target neural network 11, and output the stress triaxiality of the target material point. The first operator neural network is a deep neural network used to learn and approximate operator relationships. It can learn the mapping from one function space to another function space and solve high-dimensional problems that are difficult to handle with traditional methods. This embodiment does not limit the network structure used by the first operator neural network. For example, FNO (Fourier Neural Operator), DeepONet and other network structures can be used.

[0068] The first operator neural network 13 is used to input the spatial coordinate information of the target material point and the initial stress triaxiality output by the target neural network 11, and output the equivalent plastic strain of the target material point. The second operator neural network is also a deep neural network for learning and approximating operator relationships, which can learn the mapping from one function space to another function space and solve high-dimensional problems that are difficult to handle with traditional methods. This embodiment does not limit the network structure used by the second operator neural network. For example, FNO (Fourier Neural Operator), DeepONet and other network structures can be used.

[0069] In one embodiment, the first operator neural network 12 and the second operator neural network 13 may adopt the same network structure. For example, the first operator neural network 12 and the second operator neural network 13 may both adopt the DeepONet network structure, so that while two independent operator neural networks perform nonlinear learning from different perspectives to help the model better understand the complex patterns in the data, the consistency of the learning process can be ensured through the design of the same structure, making the model more robust.

[0070] Among them, the combination of the target neural network 11, the first operator neural network 12 and the second operator neural network 13 is used to simulate the nonlinear mapping relationship between the spatial coordinate information, stress triaxiality and equivalent plastic strain of the material point during the hypervelocity collision. Specifically, the spatial coordinate position of the target material point after the collision represents the moving distance of the target material point after the collision. The target neural network 11 can predict the changing trend of the equivalent plastic strain and stress triaxiality at the target material point through the moving distance, that is, the initial equivalent plastic strain and the initial stress triaxiality, so as to learn the relationship between the three. The first operator neural network 12 can predict the stress triaxiality of the target material point according to the changing trend and moving distance of the equivalent plastic strain of the target material point. The second operator neural network 13 can predict the equivalent plastic strain of the target material point according to the changing trend and moving distance of the stress triaxiality of the target material point. Finally, the combination of the target neural network 11, the first operator neural network 12 and the second operator neural network 13 can be used to simulate the nonlinear mapping relationship between the spatial coordinate information, stress triaxiality and equivalent plastic strain of the material point during the hypervelocity collision.

[0071] Next, Figure 1 The training process of the fusion physical information neural network 10 of the structure shown is explained.

[0072] In one embodiment, the first operator neural network 12 and the second operator neural network 13 may be trained first, and then the target neural network 11 may be trained based on the trained first operator neural network 12 and the second operator neural network 13 .

[0073] Before training begins, this embodiment needs to prepare a training data set in advance, which includes sample data of the ultra-high-speed collision process, such as the sample space coordinate information of all sample material points at each moment in the Taylor rod collision process and the corresponding equivalent plastic strain measured values ​​and stress triaxiality measured values.

[0074] The training method of the first operator neural network 12 can be to input the sample space coordinate information and the equivalent plastic strain measured value of the sample material point at any moment into the initial first operator neural network to be trained, obtain the stress triaxiality simulation value output by the initial first operator neural network, determine the first network loss according to the difference between the stress triaxiality simulation value and the stress triaxiality measured value of the target material point at that moment, adjust the network parameters of the initial first operator neural network according to the first network loss, and obtain the first operator neural network when the training end condition is met.

[0075] The training method of the second operator neural network 13 can be to input the sample space coordinate information and the actual measured value of stress triaxiality of the sample material point at any moment into the initial second operator neural network to be trained, obtain the equivalent plastic strain simulation value output by the initial second operator neural network, determine the second network loss according to the difference between the equivalent plastic strain simulation value and the actual measured value of the equivalent plastic strain of the target material point at that moment, adjust the network parameters of the initial second operator neural network according to the second network loss, and obtain the second operator neural network when the training end condition is met.

[0076] The training process is described in detail below by taking the first operator neural network 12 using the DeepONet network structure as an example.

[0077] First, let’s introduce DeepONet. The DeepONet architecture is as follows Figure 2 As shown in the figure, the DeepONet network proposed by Lu et al. can be used as the basic structure, and a branch network (Branch net) and a trunk network (Trunk net) are introduced, where the branch network is used to process the input function and the trunk network is used to process the output position, thereby achieving efficient approximation of nonlinear operators. The design goal of this network is to reduce the generalization error in training through structural innovation, and theoretically analyze the relationship between the number of sensors, the type of input function and the approximation error, so as to ensure the effectiveness and efficiency of DeepONet in operator learning. Among them, the number of sensors refers to the number of points sampled after the input function is discretized in space.

[0078] Figure 2 A shows the input and output of DeepONet, Figure 2 B shows the input and output training data, Figure 2 C shows the stacked DeepONet network structure. Figure 2 D shows the non-stacked DeepONet network structure. Define the DeepONet learning operator G:u→G(u), where the branch network is [u(x1),u(x2),...,u(x i ),...,u(x m )] as input, [b1,b2,...,b j ,...,b p ] as output, the backbone network takes y as input, [t1,t2,...,t j ,...,t p ] as output. The final output of operator learning is:

[0079]

[0080] Among them, b k and tk are the outputs of the branch network and the trunk network, respectively, and b0 is a bias term used to improve the generalization ability of the model. The combination of the branch network and the trunk network ensures that the operator neural network can capture the nonlinear relationship between the input function and the output position, and achieves the approximation of the nonlinear operator.

[0081] In the embodiment of the present disclosure, the first operator neural network 12 may adopt Figure 2 The non-stacked DeepONet network structure shown in D can extract input functions from the predefined input space to train DeepONet during the training of the initial first operator neural network. The extracted input function can be represented by the sampled value after the input function is discretized in space. For example, it can be the equivalent plastic strain measured values ​​corresponding to each of the multiple sample material points obtained by sampling, that is, the equivalent plastic strain measured value sequence. Exemplarily, the input data of the branch network can be the equivalent plastic strain measured value sequence, and the branch network extracts the input data to obtain the corresponding eigenvector, and the input data of the trunk network is the sample space coordinate information of the sample material point, and the trunk network extracts the input data to obtain the corresponding eigenvector, and finally outputs the stress triaxiality simulation value according to the obtained two eigenvectors.

[0082] The first network loss can be determined by a loss function based on the difference between the simulated value of stress triaxiality and the measured value of stress triaxiality in the training data set. The loss function is used to determine the gap between the actual output and the expected output. This embodiment does not limit the specific loss function to be used.

[0083] In a specific implementation, the network parameters in the initial first operator neural network can be adjusted by back propagation. When the network iteration end condition is reached, the network training is terminated to obtain the trained first operator neural network 12. This embodiment does not limit the preset stop training condition. Among them, the condition can be that the iteration reaches a certain number of times, or the loss value is less than a certain threshold.

[0084] In this way, this specific DeepONet architecture can predict the stress triaxiality corresponding to the material points at random positions in the "output position" by learning the nonlinear mapping relationship between the equivalent plastic strain, stress triaxiality and spatial coordinates to be solved, and can accurately predict and simulate data outside the training data.

[0085] The training method of the second operator neural network 13 may refer to the training method of the first operator neural network 12, which will not be described in detail in this embodiment.

[0086] After completing the learning and training of the first and second operator neural networks, a concept similar to transfer learning is adopted, and the pre-trained DeepONet is used as a sub-module to build a fusion physical information neural network 10 based on it, and the target neural network 11 is updated, optimized, and trained in combination with new input data.

[0087] When training the target neural network 11, the disclosed embodiment may continue to use the previous training data set or use a new training data set. Exemplarily, a complete Taylor rod collision process data may be collected, including the spatial coordinates of all material points at each moment in the process and the corresponding equivalent plastic strain actual value ε(x, y, z) and stress triaxiality actual value σ*(x, y, z), as a training data set. It should be noted that the actual values ​​of variables such as equivalent plastic strain and stress triaxiality are difficult to obtain through actual measurement, and need to be obtained through physical equation information and calculation of physical equations such as material point method.

[0088] Exemplarily, the target neural network can be a traditional neural network NN with initialization parameters W and B, and the calculation formula is y=tanh(Wx+b), where x is the input, y is the output, and both x and y are matrices; W is the weight matrix, and B is the bias term, which will be continuously updated during the training process.

[0089] The following is combined with Figure 3 The training process diagram shown in the figure illustrates the training process of the target neural network, which may include the following processing:

[0090] The sample space coordinate information of the sample material points is input into the initial neural network, and the initial values ​​of the equivalent plastic strain and the stress triaxiality output by the initial neural network are obtained.

[0091] like Figure 3 As shown in the figure, the sample space coordinate information (x, y, z) at a certain moment is used as the input of the initial neural network NN, and two variables are output: the initial value of equivalent plastic strain ε′(x, y, z) and the initial value of stress triaxiality σ * ′(x,y,z).

[0092] The sample space coordinate information and the initial value of the equivalent plastic strain are input into a pre-trained first operator neural network to obtain the stress triaxiality prediction value of the sample material point output by the first operator neural network.

[0093] For example, input (x, y, z) and ε′ Output

[0094] The sample space coordinate information and the initial value of stress triaxiality are input into the pre-trained second operator neural network to obtain the equivalent plastic strain prediction value of the sample material point output by the second operator neural network.

[0095] For example, (x, y, z) and σ *′ Enter DeepONetG ε , output

[0096] The third network loss is determined based on the difference between the measured value of stress triaxiality and the initial value of stress triaxiality, and the difference between the measured value of equivalent plastic strain and the initial value of equivalent plastic strain.

[0097] For example, according to the training data set Data * (x,y,z) and σ *’ The difference between (x, y, z), and the difference between ε(x, y, z) and ε'(x, y, z), and the third network loss (Loss of measurements) is determined by the loss function. This embodiment does not limit the loss function used. For example, MSE (MeanSquared Error, root mean square error), L2 regularization, cross entropy and other loss functions can be directly introduced from tensorflow.

[0098] The fourth network loss is determined based on the difference between the measured value of stress triaxiality and the predicted value of stress triaxiality, and the difference between the measured value of equivalent plastic strain and the predicted value of equivalent plastic strain.

[0099] For example, according to σ*(x,y,z) and The difference between ε(x,y,z) and The difference between them is determined by the loss function, and the fourth network loss (Loss of Operators) is determined by the loss function. This embodiment does not limit the loss function used. For example, MSE (Mean Squared Error), L2 regularization, cross entropy and other loss functions can be directly introduced from tensorflow.

[0100] The network parameters of the initial neural network are adjusted according to the third network loss and the fourth network loss, and when the training end condition is reached, the target neural network is obtained.

[0101] For example, the Loss of measurements and the Loss of Operators can be weighted and summed according to certain weights to obtain the Total Loss, and the network parameters in the initial neural network can be adjusted by back propagation. When the network iteration end condition is reached, the network training is terminated. This embodiment does not limit the preset stop training condition. Among them, the condition can be that the iteration reaches a certain number of times, or the loss value Total Loss is less than a certain threshold. After the training is completed, the training target neural network is obtained.

[0102] During the training process, the neural network model in the embodiment of the present disclosure uses the spatial coordinate information as the input of the target neural network, outputs the initial value of the equivalent plastic strain and the initial value of the stress triaxiality, and uses the spatial coordinate information and the initial value of the equivalent plastic strain to input the first operator neural network to learn to predict the stress triaxiality. At the same time, the spatial coordinate information and the initial value of the stress triaxiality are input into the second operator neural network to learn to predict the equivalent plastic strain. This parallel two-way prediction mechanism can enhance the adaptability of the model in complex scenarios, and can more fully allow the neural network model to learn the nonlinear mapping relationship between the variables, so that ultimately the equivalent plastic strain and stress triaxiality of the material point can be accurately simulated using only the spatial coordinate information of the material point.

[0103] In the above embodiment, the network parameters of the first and second operator neural networks are not adjusted during the training of the target neural network. In other embodiments, the network parameters of the first operator neural network and the second operator neural network may be updated and optimized during the training of the target neural network, for example, the network parameters of the target neural network, the first operator neural network, and the second operator neural network may be adjusted to different degrees according to the Total Loss.

[0104] The application process of the trained neural network model is explained below.

[0105] like Figure 4 As shown, Figure 4 is a flowchart of a method for simulating a hypervelocity collision process based on a neural network, which is shown in at least one embodiment of the present disclosure. The method can be used to simulate hypervelocity collision processes such as Taylor rod impact experiments, spherical collision experiments, and impact compression experiments. The network structure of the neural network model used can be as follows: Figure 1 As shown, the method includes the following steps. It should be noted that this embodiment does not limit the order of executing the steps.

[0106] In step 402, spatial coordinate information of a target material point at a target time is obtained, where the target material point is a material point of an object to be simulated during a hypervelocity collision.

[0107] In this step, for the hypervelocity collision process to be simulated, the spatial coordinate information of the target material point to be simulated at any target time can be obtained. The spatial coordinate information can be three-dimensional coordinate information. In other embodiments, it can also be coordinate information of other dimensions.

[0108] In step 404, the spatial coordinate information is input into a pre-trained target neural network to obtain the initial equivalent plastic strain and initial stress triaxiality output by the target neural network.

[0109] In this step, the spatial coordinate information of the target material point is input into the target neural network, and the initial equivalent plastic strain and initial stress triaxiality of the target material point preliminarily predicted by the target neural network can be obtained. The target neural network can be trained by any method of the above embodiments.

[0110] In step 406, the spatial coordinate information and the initial equivalent plastic strain are input into a pre-trained first operator neural network to obtain the stress triaxiality of the target material point at the target time output by the first operator neural network.

[0111] In this step, the spatial coordinate information of the target material point and the initial equivalent plastic strain of the target material point output in step 404 are input into the first operator neural network to obtain the three-dimensional stress of the simulated target material point at the target time. The first operator neural network can be trained by any method of the above embodiments.

[0112] In one embodiment, the first operator neural network adopts a DeepONet network structure, and the first operator neural network is composed of a first branch network and a first trunk network. In this step, the initial equivalent plastic strain can be input into the first branch network to extract the initial equivalent plastic strain eigenvector, and the spatial coordinate information can be input into the first trunk network to extract the first spatial coordinate eigenvector. According to the initial equivalent plastic strain eigenvector and the first spatial coordinate eigenvector, for example, the two eigenvectors are multiplied to obtain the stress triaxiality of the simulated target material point.

[0113] In this way, the combination of the first branch network and the first trunk network ensures that the first operator neural network can capture the nonlinear relationship between spatial coordinate information, equivalent plastic strain and stress triaxiality, realize the approximation of the nonlinear operator, and can accurately simulate and predict the stress triaxiality based on the spatial coordinate information and the initial equivalent plastic strain.

[0114] In step 408, the spatial coordinate information and the initial stress triaxiality are input into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target material point at the target time output by the second operator neural network.

[0115] In this step, the spatial coordinate information of the target material point and the initial stress triaxiality of the target material point outputted in step 404 are inputted into the second operator neural network to obtain the equivalent plastic strain of the simulated target material point at the target time. The second operator neural network can be trained by any of the methods in the above embodiments. It should be noted that the first operator neural network and the second operator neural network constitute a parallel bidirectional prediction mechanism, and step 406 and step 408 can be executed in parallel or successively.

[0116] In one embodiment, the second operator neural network adopts a DeepONet network structure, and the second operator neural network is composed of a second branch network and a second trunk network. In this step, the initial stress triaxiality can be input into the second branch network to extract the initial stress triaxiality feature vector, and the spatial coordinate information can be input into the second trunk network to extract the second spatial coordinate feature vector. According to the initial stress triaxiality feature vector and the second spatial coordinate feature vector, for example, the two feature vectors are multiplied to obtain the equivalent plastic strain of the simulated target material point.

[0117] In this way, the combination of the second branch network and the second trunk network ensures that the second operator neural network can capture the nonlinear relationship between the spatial coordinate information, equivalent plastic strain and stress triaxiality, realize the approximation of the nonlinear operator, and can accurately simulate and predict the equivalent plastic strain based on the spatial coordinate information and the initial stress triaxiality.

[0118] The combination of the target neural network, the first operator neural network and the second operator neural network is used to better simulate the nonlinear mapping relationship between the spatial coordinate information of material points, stress triaxiality and equivalent plastic strain during the hypervelocity collision process, and approximate the actual hypervelocity collision process.

[0119] The technical solution of the disclosed embodiment provides a method for simulating ultra-high-speed collision processes based on a neural network. The neural network model obtained by pre-training through an operator learning method based on deep learning adopts a parallel two-way prediction mechanism, which can enhance the adaptability of the model in complex scenarios. It can learn the changing trends of various variables when the material is deformed during the collision process under limited measurement data. Therefore, when simulating the ultra-high-speed collision process, the stress triaxiality and equivalent plastic strain of the target material point can be accurately and quickly simulated based on the spatial coordinates of the target material point, overcoming the limitations of traditional numerical methods in large deformation and high nonlinear problems, effectively improving the simulation accuracy of ultra-high-speed collision processes such as Taylor bar impact, providing a more efficient solution for engineering simulation, and helping to promote research and application development in related fields.

[0120] Take Taylor rod collision as an example, Figure 5As shown, after completing the construction of the overall neural network model, the relevant parameters in the model are preliminarily adjusted, and then the model is trained and learned through the collected Taylor rod collision process data set to obtain a Taylor rod collision process simulation model.

[0121] The results of the trained model simulating the Taylor rod collision process are as follows Figure 6 shown. Figure 6 The test results of the equivalent plastic strain (first row) and stress triaxiality (second row) predicted by the neural network model at the 500th time step after the collision are shown respectively, where the three-dimensional coordinate axes in the figure represent the spatial distribution of material points and the colors represent the numerical values.

[0122] Taking the equivalent plastic strain results of the first row as an example, the first picture is the model predicting the equivalent plastic strain to obtain the equivalent plastic strain prediction value, and the color indicates the size of the prediction value; the second picture is the equivalent plastic strain measured value, and the color indicates the size of the measured value; the third picture is the error value between the equivalent plastic strain measured value and the predicted value, and the color indicates the size of the error value. It can be seen that the error value is almost 0 (indicated by white). Similarly, the error value between the stress triaxiality measured value and the predicted value in the second row is also mostly 0. The neural network model proposed in the embodiment of the present disclosure has the ability to accurately simulate and predict in the Taylor rod collision problem.

[0123] The neural network-based hypervelocity collision process simulation method proposed in the embodiment of the present disclosure can effectively simulate complex physical problems (such as Taylor rod collision) by combining limited measurement data with calculation methods. Figure 6 The Taylor bar collision process can be accurately simulated, and because the equivalent plastic sample and stress triaxiality can be quickly predicted based on the spatial coordinate information after the model is trained, without the need for complex calculations, the simulation speed is increased by 10 times compared with the traditional numerical method.

[0124] like Figure 7 As shown, Figure 7 is a block diagram of a hypervelocity collision process simulation device based on a neural network according to at least one embodiment of the present disclosure, the device comprising:

[0125] A data acquisition module 71, used to acquire spatial coordinate information of a target material point at a target time, wherein the target material point is a material point of an object to be simulated during a hypervelocity collision;

[0126] The target network module 72 is used to input the spatial coordinate information into the pre-trained target neural network to obtain the initial equivalent plastic strain and initial stress triaxiality output by the target neural network;

[0127] The first operator module 73 is used to input the spatial coordinate information and the initial equivalent plastic strain into a pre-trained first operator neural network to obtain the stress triaxiality of the target material point at the target time output by the first operator neural network;

[0128] The second operator module 74 is used to input the spatial coordinate information and the initial stress triaxiality into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target material point at the target time output by the second operator neural network;

[0129] The combination of the target neural network, the first operator neural network and the second operator neural network is used to simulate the nonlinear mapping relationship between the spatial coordinate information of material points, stress triaxiality and equivalent plastic strain during a hypervelocity collision.

[0130] In some embodiments, the first operator neural network and the second operator neural network adopt the same network structure.

[0131] In some embodiments, the first operator neural network adopts a deep operator network DeepONet network structure; the first operator neural network is composed of a first branch network and a first trunk network; the first operator module 73, when used to input spatial coordinate information and initial equivalent plastic strain into the pre-trained first operator neural network to obtain the stress triaxiality of the target material point output by the first operator neural network at the target time, is specifically used to:

[0132] The initial equivalent plastic strain is input into the first branch network to extract the initial equivalent plastic strain eigenvector; the spatial coordinate information is input into the first trunk network to extract the first spatial coordinate eigenvector; based on the initial equivalent plastic strain eigenvector and the first spatial coordinate eigenvector, the stress triaxiality of the simulated target material point at the target time is obtained.

[0133] In some embodiments, the second operator neural network adopts a DeepONet network structure; the second operator neural network is composed of a second branch network and a second trunk network; the second operator module 74, when used to input the spatial coordinate information and the initial stress triaxiality into the pre-trained second operator neural network, obtains the equivalent plastic strain of the target material point output by the second operator neural network at the target time, is specifically used to:

[0134] The initial stress triaxiality is input into the second branch network to extract the initial stress triaxiality eigenvector; the spatial coordinate information is input into the second trunk network to extract the second spatial coordinate eigenvector; based on the initial stress triaxiality eigenvector and the second spatial coordinate eigenvector, the equivalent plastic strain of the simulated target material point at the target time is obtained.

[0135] In some embodiments, the first operator neural network is trained in the following manner:

[0136] The sample space coordinate information of the sample material points and the measured value of the equivalent plastic strain are input into the initial first operator neural network to be trained, and the stress triaxiality simulation value output by the initial first operator neural network is obtained; the first network loss is determined according to the difference between the stress triaxiality simulation value and the measured value of the stress triaxiality; the network parameters of the initial first operator neural network are adjusted according to the first network loss, and the first operator neural network is obtained when the training end condition is met.

[0137] In some embodiments, the second operator neural network is trained in the following manner: the sample space coordinate information of the sample material points and the measured values ​​of stress triaxiality are input into the initial second operator neural network to be trained to obtain the equivalent plastic strain simulation value output by the initial second operator neural network; the second network loss is determined based on the difference between the equivalent plastic strain simulation value and the measured value of the equivalent plastic strain; the network parameters of the initial second operator neural network are adjusted based on the second network loss, and the second operator neural network is obtained when the training end conditions are met.

[0138] In some embodiments, the target neural network is trained in the following manner: the sample space coordinate information of the sample material point is input into the initial neural network to obtain the initial value of the equivalent plastic strain and the initial value of the stress triaxiality output by the initial neural network; the sample space coordinate information and the initial value of the equivalent plastic strain are input into a pre-trained first operator neural network to obtain the stress triaxiality prediction value of the sample material point output by the first operator neural network; the sample space coordinate information and the initial value of the stress triaxiality are input into a pre-trained second operator neural network to obtain the equivalent plastic strain prediction value of the sample material point output by the second operator neural network; the third network loss is determined based on the difference between the measured value of the stress triaxiality and the initial value of the stress triaxiality, and the difference between the measured value of the equivalent plastic strain and the initial value of the equivalent plastic strain; the fourth network loss is determined based on the difference between the measured value of the stress triaxiality and the predicted value of the stress triaxiality, and the difference between the measured value of the equivalent plastic strain and the predicted value of the equivalent plastic strain; the network parameters of the initial neural network are adjusted based on the third network loss and the fourth network loss, and the target neural network is obtained when the training end condition is met.

[0139] It should be noted that the training of the first operator neural network can be performed on the above-mentioned ultra-high-speed collision process simulation device based on the neural network, or on other devices; the training of the second operator neural network can be performed on the above-mentioned ultra-high-speed collision process simulation device based on the neural network, or on other devices; the training of the target neural network can be performed on the above-mentioned ultra-high-speed collision process simulation device based on the neural network, or on other devices.

[0140] The implementation process of the functions and effects of each module in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, which will not be repeated here.

[0141] The present disclosure also provides an electronic device, such as Figure 8 As shown, the electronic device includes a memory 81 and a processor 82. The memory 81 is used to store computer instructions that can be executed on the processor, and the processor 82 is used to implement the neural network-based ultra-high-speed collision process simulation method of any embodiment of the present disclosure when executing the computer instructions.

[0142] The embodiments of the present disclosure also provide a computer program product, which includes a computer program / instructions. When the computer program / instructions are executed by a processor, the neural network-based hypervelocity collision process simulation method of any embodiment of the present disclosure is implemented.

[0143] The embodiments of the present disclosure also provide a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the method for simulating a hypervelocity collision process based on a neural network according to any embodiment of the present disclosure is implemented.

[0144] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this specification. Ordinary technicians in this field can understand and implement it without paying creative work.

[0145] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0146] Those skilled in the art will readily appreciate other embodiments of the specification after considering the specification and practicing the invention claimed herein. The specification is intended to cover any variations, uses or adaptations of the specification that follow the general principles of the specification and include common knowledge or customary techniques in the art that are not claimed in the specification. The specification and examples are to be considered exemplary only, and the true scope and spirit of the specification are indicated by the following claims.

[0147] It should be understood that the present description is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present description is limited only by the appended claims.

[0148] The above description is only a preferred embodiment of this specification and is not intended to limit this specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this specification should be included in the scope of protection of this specification.

Claims

1. A method for simulating a hypervelocity collision process based on a neural network, characterized in that: The method comprises: Acquire spatial coordinate information of a target material point at a target time, wherein the target material point is a material point of an object to be simulated during a hypervelocity collision; Inputting the spatial coordinate information into a pre-trained target neural network to obtain the initial equivalent plastic strain and initial stress triaxiality output by the target neural network; Inputting the spatial coordinate information and the initial equivalent plastic strain into a pre-trained first operator neural network to obtain the stress triaxiality of the target material point at the target time output by the first operator neural network; Inputting the spatial coordinate information and the initial stress triaxiality into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target material point at the target time output by the second operator neural network; The combination of the target neural network, the first operator neural network and the second operator neural network is used to simulate the nonlinear mapping relationship between the spatial coordinate information of material points, stress triaxiality and equivalent plastic strain during a hypervelocity collision.

2. The method according to claim 1, characterized in that The first operator neural network and the second operator neural network adopt the same network structure.

3. The method according to claim 1 or 2, characterized in that: The first operator neural network adopts a deep operator network DeepONet network structure; the first operator neural network consists of a first branch network and a first trunk network; The step of inputting the spatial coordinate information and the initial equivalent plastic strain into a pre-trained first operator neural network to obtain the stress triaxiality of the target material point at the target time output by the first operator neural network includes: Inputting the initial equivalent plastic strain into the first branch network to extract an initial equivalent plastic strain eigenvector; Inputting the spatial coordinate information into a first backbone network to extract a first spatial coordinate feature vector; The simulated stress triaxiality of the target material point at the target time is obtained according to the initial equivalent plastic strain eigenvector and the first space coordinate eigenvector.

4. The method according to claim 1 or 2, characterized in that: The second operator neural network adopts a DeepONet network structure; the second operator neural network consists of a second branch network and a second trunk network; The step of inputting the spatial coordinate information and the initial stress triaxiality into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target material point at the target time output by the second operator neural network includes: Inputting the initial stress triaxiality into the second branch network to extract the initial stress triaxiality feature vector; Inputting the spatial coordinate information into a second backbone network to extract a second spatial coordinate feature vector; According to the initial stress triaxiality eigenvector and the second spatial coordinate eigenvector, the simulated equivalent plastic strain of the target material point at the target time is obtained.

5. The method according to claim 1, characterized in that The first operator neural network is trained in the following manner: Inputting the sample space coordinate information and the equivalent plastic strain measured value of the sample material point into the initial first operator neural network to be trained, and obtaining the stress triaxiality simulation value output by the initial first operator neural network; determining a first network loss according to a difference between the stress triaxiality simulation value and the stress triaxiality measured value; The network parameters of the initial first operator neural network are adjusted according to the first network loss, and when the training end condition is reached, the first operator neural network is obtained.

6. The method according to claim 1, characterized in that The second operator neural network is trained in the following manner: Inputting the sample space coordinate information of the sample material point and the measured value of stress triaxiality into the initial second operator neural network to be trained, and obtaining the equivalent plastic strain simulation value output by the initial second operator neural network; Determining a second network loss according to a difference between the equivalent plastic strain simulation value and the equivalent plastic strain measured value; The network parameters of the initial second operator neural network are adjusted according to the second network loss, and when the training end condition is reached, the second operator neural network is obtained.

7. The method according to claim 1, characterized in that The target neural network is trained in the following way: Inputting sample space coordinate information of sample material points into an initial neural network to obtain an initial value of equivalent plastic strain and an initial value of stress triaxiality output by the initial neural network; Inputting the sample space coordinate information and the initial value of the equivalent plastic strain into a pre-trained first operator neural network to obtain a stress triaxiality prediction value of the sample material point output by the first operator neural network; Inputting the sample space coordinate information and the initial value of the stress triaxiality into a pre-trained second operator neural network to obtain an equivalent plastic strain prediction value of the sample material point output by the second operator neural network; Determine a third network loss according to a difference between the measured value of the stress triaxiality and the initial value of the stress triaxiality, and a difference between the measured value of the equivalent plastic strain and the initial value of the equivalent plastic strain; Determine a fourth network loss according to a difference between the measured value of the stress triaxiality and the predicted value of the stress triaxiality, and a difference between the measured value of the equivalent plastic strain and the predicted value of the equivalent plastic strain; The network parameters of the initial neural network are adjusted according to the third network loss and the fourth network loss, and when the training end condition is reached, the target neural network is obtained.

8. A device for simulating a hypervelocity collision process based on a neural network, characterized in that: The device comprises: A data acquisition module, used to acquire spatial coordinate information of a target material point at a target time, wherein the target material point is a material point of an object to be simulated during a hypervelocity collision; A target network module, used for inputting the spatial coordinate information into a pre-trained target neural network to obtain the initial equivalent plastic strain and initial stress triaxiality output by the target neural network; A first operator module, used for inputting the spatial coordinate information and the initial equivalent plastic strain into a pre-trained first operator neural network, and obtaining the stress triaxiality of the target material point at the target time output by the first operator neural network; A second operator module is used to input the spatial coordinate information and the initial stress triaxiality into a pre-trained second operator neural network to obtain the equivalent plastic strain of the target material point at the target time output by the second operator neural network; The combination of the target neural network, the first operator neural network and the second operator neural network is used to simulate the nonlinear mapping relationship between the spatial coordinate information of material points, stress triaxiality and equivalent plastic strain during a hypervelocity collision.

9. An electronic device, characterized in that: The device includes a memory and a processor, wherein the memory is used to store computer instructions that can be executed on the processor, and the processor is used to implement the neural network-based hypervelocity collision process simulation method described in any one of claims 1 to 7 when executing the computer instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the neural network-based hypervelocity collision process simulation method according to any one of claims 1 to 7 is implemented.

Images