A method for intelligent design of a random short fiber composite structure

By modeling the fiber filling process as a reinforcement learning model and combining it with a prediction model for rapid prediction of equivalent performance, the problems of computational efficiency and design accuracy in composite material structure design are solved, and efficient and automated intelligent structural design is realized.

CN122433419APending Publication Date: 2026-07-21BEIJING INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610709356.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-21
Publication Date
2026-07-21

AI Technical Summary

Technical Problem

Existing composite material structure design methods struggle to balance computational efficiency and design accuracy. Random adsorption methods have low modeling efficiency, finite element simulation has high computational costs, and there is a lack of unified sample library management and geometric modeling decision-making, making it difficult to achieve efficient and closed-loop intelligent structural design.

Method used

The fiber filling process is modeled as a reinforcement learning model, and the equivalent performance is quickly predicted by combining it with the prediction model. Through parameter feedback iteration, the representative volume element model of the composite material automatically converges to the given range of mechanical properties of the fiber and matrix, replacing the traditional stochastic adsorption modeling and finite element simulation process.

Benefits of technology

It has achieved breakthroughs in design automation, computational cost control, and engineering applicability, and can balance computational efficiency and design accuracy, thereby improving the automation level and computational efficiency of the design.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122433419A_ABST
    Figure CN122433419A_ABST
Patent Text Reader

Abstract

The application discloses a kind of random short fiber composite structure intelligent design methods, belong to learning field.Method includes: defining fiber aspect ratio range in reinforcement learning model, defining mechanical property range in prediction model;Set target performance, find the representative volume element sample closest to target performance in prediction experience pool, compare the sample with target performance, adjust fiber aspect ratio input to reinforcement learning model;Reinforcement learning model selects the position and angle corresponding to the maximum action value mapping function to fill the current fiber, to obtain new representative volume element sample;Mechanical property of fiber and matrix is adjusted to new representative volume element sample using prediction model, compare the performance result of new representative volume element sample with target performance, carry out new round of iteration, until the representative volume element model output that meets target performance is obtained.This scheme can give consideration to calculation efficiency and design accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of learning technology, and in particular to an intelligent design method for random short fiber composite material structures. Background Technology

[0002] Short fiber reinforced composites are widely used in aerospace, automotive lightweighting, and civil engineering due to their advantages such as high specific strength, high specific stiffness, and flexible molding processes. For these materials, the macroscopic equivalent properties significantly depend on the fiber volume fraction, aspect ratio, orientation distribution, and spatial arrangement characteristics at the microscopic level. In engineering, representative volume elements (RVEs) are typically constructed. Given the number of fibers and geometric parameters, a fiber-matrix geometric model is generated using the Random Sequential Adsorption (RSA) method. This model is then combined with the Finite Element Method (FEM) to apply periodic boundary conditions or equivalent loads to predict parameters such as the effective elastic modulus and shear modulus of the material. This forms a micromechanical analysis process of "RSA modeling + FEM simulation," which is used to guide material formulation and structural design.

[0003] Current composite material structure design methods still primarily rely on "RSA modeling + FEM simulation." Even with the introduction of learning and reinforcement learning, these methods mostly involve local substitution at a specific stage of modeling or prediction, making it difficult to construct a highly efficient and closed-loop intelligent structural design process as a whole. On the one hand, when the number of fibers, volume fraction, or aspect ratio is large, the random adsorption method requires extensive overlap checks and rejection operations in the later stages of the modeling process, significantly reducing efficiency and hindering multi-round iterations and large-scale sample generation. On the other hand, finite element simulation is computationally expensive under fine meshes and multi-parameter conditions, making it difficult to support rapid iterative design based on target performance requirements. Existing work based on convolutional neural networks (CNN) or reinforcement learning (DQN) largely remains at the level of offline prediction or local acceleration, lacking a unified intelligent structural design framework that combines sample library management, geometric modeling decision-making, and rapid performance identification, making it difficult to balance design accuracy, computational efficiency, and engineering usability.

[0004] Therefore, there is an urgent need to provide an intelligent design method for random short fiber composite material structures. Summary of the Invention

[0005] To address the challenge of balancing computational efficiency and design accuracy in traditional composite material structure design methods, this invention provides an intelligent design method for random short fiber composite material structures.

[0006] On the one hand, a smart design method for random short fiber composite material structures is provided, the method comprising: In the reinforcement learning model used for modeling representative volumetric units, the fiber aspect ratio range is defined; in the prediction model used for performance prediction of representative volumetric units, the mechanical property range of the fiber and the matrix is ​​defined. A target performance of the random short fiber composite material is set, and a representative volumetric unit sample that is closest to the target performance is found in the prediction experience pool. The representative volumetric unit sample is compared with the target performance to adjust the fiber aspect ratio and input it into the reinforcement learning model. The initial batch of samples in the prediction experience pool is obtained by combining the random adsorption method and the finite element method. Each sample is stored as a quaternion, which consists of the image name of the representative volumetric unit and the elastic modulus performance. The reinforcement learning model is based on a predefined Markov quadruple and calculates the action value mapping function value when each fiber is filled at different positions and angles. The position and angle corresponding to the maximum action value mapping function are selected to fill the current fiber, thus obtaining a new representative volume unit sample. The mechanical properties of the fiber and matrix of the new representative volumetric unit sample are adjusted using the prediction model to obtain the performance results of the new representative volumetric unit sample. The performance results of the new representative volumetric unit samples are compared with the target performance to adjust the fiber aspect ratio input to the reinforcement learning model for a new round of iteration until a representative volumetric unit model output that meets the target performance is obtained.

[0007] On the other hand, a smart design device for random short fiber composite material structures is provided for implementing the steps described in any method embodiment of the specification, the device comprising: Define the fiber aspect ratio range in the reinforcement learning model used for modeling representative volumetric elements, and define the mechanical property range of the fiber and matrix in the prediction model used for performance prediction of representative volumetric elements. The adjustment module is used to set the target performance of the random short fiber composite material, find the representative volumetric unit sample that is closest to the target performance in the prediction experience pool, compare the representative volumetric unit sample with the target performance, and adjust the fiber aspect ratio input to the reinforcement learning model. The initial batch of samples in the prediction experience pool is obtained by combining the random adsorption method and the finite element method. Each sample is stored as a quaternion, which consists of the image name of the representative volumetric unit and the elastic modulus performance. The modeling module, based on a predefined Markov quadruple, calculates the action value mapping function value for each fiber filled at different positions and angles, and selects the position and angle corresponding to the maximum action value mapping function to fill the current fiber, thereby obtaining a new representative volume unit sample. The prediction module is used to adjust the mechanical properties of the fiber and matrix of the new representative volumetric unit sample using the prediction model, so as to obtain the performance results of the new representative volumetric unit sample. The iteration module compares the performance results of the new representative volumetric unit samples with the target performance to adjust the fiber aspect ratio input to the reinforcement learning model and perform a new round of iteration until a representative volumetric unit model output that meets the target performance is obtained.

[0008] On the other hand, a computer device is provided, the computer device including a memory and a processor, the memory for storing a computer program, and the processor for executing the computer program stored in the memory to implement the steps of the method described above.

[0009] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the steps of the method described above.

[0010] On the other hand, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0011] The technical solution provided by this invention can bring at least the following beneficial effects: The fiber filling process is modeled as a reinforcement learning model, and combined with a prediction model, the equivalent performance of the RVE image is rapidly predicted, replacing the traditional stochastic adsorption modeling and finite element simulation process. Through parameter feedback iteration, the representative volume element model of the composite material can automatically converge to the given mechanical property range of the fiber and matrix, obtaining a representative volume element model that meets the target performance. Compared with the traditional design process that relies on manual parameter tuning and repeated simulations, it has more outstanding application value and promotion prospects in terms of design automation, computational cost control, and engineering applicability, and can balance computational efficiency and design accuracy. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of a method for intelligent design of random short fiber composite material structures according to an embodiment of the present invention; Figure 2This is a diagram of an intelligent design model architecture for a random short fiber composite material structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the training process of a reinforcement learning model provided in an embodiment of the present invention; Figure 4 This is an iterative process diagram under two strategies provided in an embodiment of the present invention; Figure 5 This is a representative volumetric unit model image that conforms to the target performance obtained by iterative analysis of the properties of each phase material according to an embodiment of the present invention. Figure 6 This is a representative volumetric unit model image that meets the target performance, obtained by iterating the fiber geometric parameters according to an embodiment of the present invention. Figure 7 This is a structural diagram of a random short fiber composite material structure intelligent design device provided in an embodiment of the present invention; Figure 8 This is a hardware architecture diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0015] The specific implementation of the above concept is described below.

[0016] Please refer to Figure 1 The present invention provides a smart design method for random short fiber composite material structures, the method comprising: Step 100: Define the fiber aspect ratio range in the reinforcement learning model used for modeling representative volumetric elements, and define the mechanical property range of the fiber and matrix in the prediction model used for performance prediction of representative volumetric elements. Step 102: Set the target performance of the random short fiber composite material, find the representative volumetric unit sample that is closest to the target performance in the prediction experience pool, compare the representative volumetric unit sample with the target performance, and adjust the fiber aspect ratio input to the reinforcement learning model; the initial batch of samples in the prediction experience pool is obtained by combining the random adsorption method and the finite element method, and each sample is stored as a quaternion, which consists of the image name of the representative volumetric unit and the elastic modulus performance; Step 104: The reinforcement learning model calculates the action value mapping function value for each fiber filled at different positions and angles based on the predefined Markov quadruples, and selects the position and angle corresponding to the maximum action value mapping function to fill the current fiber, thereby obtaining a new representative volume cell sample. Step 106: Use the prediction model to adjust the mechanical properties of the fiber and matrix of the new representative volumetric unit sample to obtain the performance results of the new representative volumetric unit sample; Step 108: Compare the performance results of the new representative volumetric unit sample with the target performance to adjust the fiber aspect ratio input to the reinforcement learning model and perform a new round of iteration until a representative volumetric unit model output that meets the target performance is obtained.

[0017] In this embodiment of the invention, the fiber filling process is modeled as a reinforcement learning model, and combined with a prediction model to quickly predict the equivalent performance of the RVE image, replacing the traditional stochastic adsorption modeling and finite element simulation process. Through parameter feedback iteration, the representative volume element model of the composite material can automatically converge to the given mechanical property range of the fiber and matrix, obtaining a representative volume element model that meets the target performance. Compared with the traditional design process that relies on manual parameter tuning and repeated simulations, it has more outstanding application value and promotion prospects in terms of design automation, computational cost control, and engineering applicability, and can balance computational efficiency and design accuracy.

[0018] The following description Figure 1 The execution method for each step is shown.

[0019] For step 100: In this embodiment, the mechanical properties range of the fiber and the matrix include the range of fiber elastic modulus and Poisson's ratio, and the range of matrix elastic modulus and Poisson's ratio.

[0020] Regarding step 102: You can refer to this. Figure 2In this embodiment, the sample library module mainly includes two sample pools: a prediction resource pool and a reinforcement learning experience pool. The reinforcement learning experience pool uses a first-in, first-out (FIFO) queue structure. Data is automatically generated based on the reinforcement learning model, and each sample is stored as a quaternion D(s,a,r,s'). During later experience exploration, the latest memory samples overwrite older ones, automatically updating the samples. The initial batch of samples in the prediction resource pool is obtained by combining the random adsorption method and the finite element method. Each sample is stored as a quaternion T(imgName, E1, E2, G12), where imgName represents the name of the representative volumetric unit image, and E1, E2, and G12 are the elastic modulus in the first direction, the elastic modulus in the second direction, and the shear modulus in the 12th direction, respectively. Given a target performance requirement, the system first searches the prediction resource pool. If a sample meeting the performance requirement is found, the image is directly output; otherwise, the sample closest to the target performance is found, the error is calculated, and feedback iteration is performed. During the feedback iteration process, samples that do not meet the requirements are not discarded, but are sent to the prediction resource pool as resource accumulation to expand the sample library capacity and increase the probability of subsequent samples meeting the design requirements.

[0021] In some implementations, representative volumetric unit samples are compared with target performance to adjust the fiber aspect ratio input to the reinforcement learning model, including: In the formula, The aspect ratio of the fiber. and The performance results for representative volumetric unit samples are the elastic modulus in the first direction and the elastic modulus in the second direction, respectively. and To represent the target performance, the target elastic modulus in the first direction and the target elastic modulus in the second direction are respectively. For fiber angle.

[0022] It should be noted that there are many designable variables in the intelligent design of random short fiber composite structures, and one or more can be selected as design variables according to the actual situation. These include material parameters of each phase (mechanical properties of the fiber and matrix, i.e., elastic modulus, Poisson's ratio) and fiber geometric parameters (aspect ratio, angle distribution, etc.). Since the uncertainty of fiber position has little impact on the overall material performance, this invention does not consider the design of position parameters. Because both material parameters and geometric parameters affect the overall material performance, considering all designable variables would result in multiple solutions that meet the requirements, and many different feedback design schemes. To simplify the design process, a feedback iteration process is designed with material parameters of each phase and fiber geometric parameters as design variables. The above formula represents the iterative strategy for adjusting the fiber aspect ratio. The geometric parameter iteration strategy adopts a collaborative iteration method of fiber aspect ratio and angle. Let the performance requirements be the elastic modulus E1 and E2 of the material in two directions, with fiber aspect ratio and angle as independent design variables, and material parameters of each phase as dependent variables. Another approach is the material parameter iteration strategy. The elastic modulus of both the fiber phase and the matrix phase will affect the overall material properties, and the matrix material properties will have a greater impact on the overall prediction results. The material parameter iteration strategy is as follows: if the prediction results differ significantly from the target properties (>5%), the elastic modulus Em of the matrix phase is iteratively adjusted; otherwise, the elastic modulus Ef of the fiber phase is iteratively adjusted.

[0023] Regarding step 104: Most current finite element geometric modeling methods for composite materials mainly employ the random adsorption method. This method, based on random filling, requires continuous evaluation of whether the newly added fibers meet the requirements, resulting in a very slow process in the later stages of modeling and severely impacting the efficiency of composite structure analysis. This invention aims to introduce reinforcement learning into the traditional RSA modeling logic, training a reinforcement learning model to assist the fiber decision-making process. The input of the reinforcement learning model is the fiber aspect ratio, and the output is the fiber RVE model image, which is then fed into a prediction model for performance prediction.

[0024] In this embodiment, the volume parameters of a representative volume unit in the reinforcement learning model are defined as side length, fiber width, and number of fibers.

[0025] It's important to note that traditional reinforcement learning methods (Q-learning, Sarsa) store Q-values ​​using a Q-matrix. Their state spaces are discrete and exhaustively enumerable. For problems with large state or action spaces, the inability to construct a sufficiently large Q-matrix prevents a complete representation of the state and action spaces. To enable reinforcement learning algorithms to handle inexhaustible action and state spaces, a non-linear neural network is used to model the Q-matrix. The Decentralized Q-N (DQN) reinforcement learning model utilizes a convolutional neural network as a non-linear mapping function Q(s) that maps action values ​​to states, thus modeling the Q-matrix and achieving a fully end-to-end learning process.

[0026] It is understandable that once a reinforcement learning model is trained, without the need for a convolutional neural network, the action value mapping function value can be calculated for each fiber placed at different positions and angles. The position and angle corresponding to the maximum action value mapping function can then be selected to fill the current fiber until the required number of fibers are filled, thus obtaining a representative volumetric unit model of the random short fiber composite material.

[0027] Next, we will explain how to train the reinforcement learning model.

[0028] You can refer to this. Figure 3 In some implementations, the reinforcement learning model is trained as shown in steps S1-S7 below: S1 initializes the parameters of the reinforcement learning network and the convolutional neural network; S2 generates an initial state sample with 1 fiber. For each initial state sample, the following is executed: S3, using a predefined action space, fill the current initial state sample with the next fiber at different positions and angles to generate several state samples with a fiber count of 2; S4, input all state samples into the convolutional neural network and the reinforcement learning network respectively; S5, the convolutional neural network extracts features from the state samples and determines whether there are fiber crossings in the input state samples. The reward values ​​of each state sample are then input into the reinforcement learning network. S6, the reinforcement learning network calculates the action value mapping function value based on each state sample, calculates the corresponding action value true value using the reward value of each state sample input by the convolutional neural network, and calculates the loss function using the action value true value and the action value mapping function value to adjust the network parameters of the reinforcement learning network; S7, based on probability coefficients, selects from several state samples with 2 fibers to put into the experience pool until all initial state samples are filled and trained; S8: For each state sample with 2 fibers in the experience pool, the next fiber at different positions and angles is filled into the state sample using a predefined action space to generate several state samples with 3 fibers. Then, the process jumps to input all state samples into the convolutional neural network and the reinforcement learning network respectively until the required number of fibers is reached, thus completing this round of training for the reinforcement learning network.

[0029] In this embodiment, the Markov quadruple is (S, A, P, R), where S is the set of all possible states s, i.e., the fiber distribution state diagram after each fiber is added. A is the spatial set of all actions a that the agent can take: a∈A, where action a is the position and angle selected during fiber addition. P is the state transition probability, i.e., the probability of the agent reaching another state after taking action a in one state, a random factor for action selection, adjusted according to the network training process; where for The next state reached after taking an action; R is the reward function, which is the immediate reward obtained by the model after taking an action, denoted as r(s,a). It needs to be tested during training. It can be assumed that when the newly added fiber meets the requirements, r>0, otherwise r<0.

[0030] In this embodiment of the invention, the action space is represented by the following formula: In the formula, For the action space, This represents the filling position of the fiber center point in the state sample. Let the angle of the fiber in the state sample be denoted as . This represents the side length of a three-dimensional representative volume element model.

[0031] In this embodiment, Discretize the fiber into N actions, each representing a specific fiber position and orientation parameter. Let the number of fiber positions be locationNum, and the orientation be discretized from 0 to 2π as orientationNum. Then N is calculated as follows: N = locationNum × orientationNum.

[0032] In some implementations, when the state sample represents a three-dimensional representative volume element model, the pixel grayscale values ​​in the state sample characterize the orientation information of the three-dimensional representative volume element model, and the correspondence between the pixel grayscale values ​​and the orientation information in the state sample is as follows: In the formula, Midpoint of the state sample pixel grayscale values, This represents the Z-coordinate value of the point corresponding to the representative three-dimensional volume element model. The directional side length of the representative three-dimensional volume element model. Here is the Z-coordinate value of the first endpoint of the fiber in the three-dimensional representative volume element model. Midpoint of the state sample The distance to the first end of the fiber. Euler angles; In this embodiment, step S5, determining whether there are fiber crossings in the input state sample, includes: Based on the grayscale value of each pixel in each fiber in the state sample, calculate the position of each pixel in the three-dimensional representative volume unit model; Calculate whether the distance between the centerlines of any two fibers is greater than the fiber width; If so, confirm that the two fibers do not cross each other; If not, it is determined that the two fibers intersect.

[0033] In this embodiment, the basic condition for fiber distribution during the filling process is that they cannot cross each other. Therefore, when the new fiber filled by action a does not cross with other fibers, a positive reward is given; when the new fiber filled by action a crosses with other fibers, a negative reward is given.

[0034] Therefore, the reward value for each state sample is calculated using the following formula: In the formula, For filling action Formed state samples The corresponding reward value, This indicates that the fibers are effective, meaning there is no cross-linking between the fibers. This indicates that the fiber is invalid, meaning that the fiber crosses. This indicates that the fiber filling has reached the required number of fibers, and the modeling ends; v>>t>0, u<0.

[0035] In this embodiment, Figure 3 The following is a schematic diagram of the reinforcement learning model training in this embodiment. The input to the convolutional neural network (CNN) is a state sample after adding new fibers using different actions. Features are extracted from the state samples, and it is determined whether fiber crossings exist in the input state samples. The reward value for each state sample is then calculated. Input into the reinforcement learning network.

[0036] The reinforcement learning network is equivalent to a non - linear mapping function Q(s) that maps state to action value. At this time, the input is no longer discrete actions, but state samples after adding new fibers using different actions. Through the processes of convolution, pooling, non - linearity, and fully - connected layers, image features are extracted layer by layer and regressed to the action value mapping function values of each discrete action [Q(s,a1), Q(s,a2), Q(s,a3)……Q(s,aN)]. During the training process, first, the reward values of each state sample input by the convolutional neural network are used to calculate the corresponding true action values, making the true action values equal to the action value mapping function values, to iteratively train the network parameters of the reinforcement learning network. The formula is: In the formula, is the true action value, is the reward value of the state sample input by the convolutional neural network, is the benefit weight, is the maximum action value mapping function value among all actions, is the action value mapping function value calculated by the reinforcement learning network.

[0037] It can be understood that after training the reinforcement learning model, the action corresponding to the largest Q is selected as the action to be executed by the agent.

[0038] In some embodiments, step S7 "selecting and putting into the experience pool several state samples with a fiber number of 2 based on the probability coefficient" includes: Obtaining the current probability coefficient and the random number generated before filling the first fiber number of the current initial state sample; where the probability coefficient gradually decays with the number of training times; Judging whether the random number is less than the current probability coefficient; If so, all state samples with a fiber number of 2 generated by filling based on the current initial state sample are put into the reinforcement learning experience pool; If not, the state sample with the largest action value mapping function value among several state samples with a fiber number of 2 generated by filling based on the current initial state sample is put into the reinforcement learning experience pool.

[0039] In this embodiment, before each action selection, a random number h between 0 and 1 is generated. If h < epsilon, the state samples of all random actions are selected, indicating that the agent is in the exploration process; if h > epsilon, the state sample with the largest action value mapping function value is selected, indicating that it is put into the experience pool according to experience. The probability coefficient epsilon is initialized to 1 and gradually decays with the number of training times. The decay formula of the probability coefficient epsilon is: Where epochNum represents the number of training iterations.

[0040] In this embodiment, the probability of epsilon determines whether to output an action based on the model's Q-value or to store random action samples in the experience pool to increase exploration opportunities. Epsilon is a decaying number, representing the agent's increasing reliance on earlier experience judgments in later stages of exploration, which can accelerate the convergence speed of the reinforcement learning network. When the experience pool reaches saturation, a first-in, first-out (FIFO) queue is used, meaning later memories overwrite earlier ones.

[0041] In some implementations, when the number of samples in the reinforcement learning experience pool reaches a set value, state samples are extracted from the reinforcement learning experience pool for memory replay training of the reinforcement learning network.

[0042] In this embodiment, after obtaining a reinforcement learning experience pool of a certain size, a small batch of data is randomly selected from the reinforcement learning experience pool as training samples.

[0043] The loss function is calculated using the following formula: In the formula, For the true value of the corresponding action, To reinforce the action-value mapping function value of the learning network, the memory playback process is a step-by-step approximation process, making the predicted value gradually approach the true value, thus making the estimate more accurate.

[0044] It is understandable that a trained reinforcement learning model is used to calculate the action value mapping function value when each fiber is filled at different positions and angles. The position and angle corresponding to the maximum action value mapping function are selected to fill the current fiber. By gradually filling the fiber to reach the required number of fibers, a representative volumetric unit model of random short fiber composite material without fiber intersections is obtained.

[0045] Regarding step 106: In this embodiment, another major obstacle affecting the efficiency of material structure design is the finite element simulation solution process. Solving large meshes is typically time-consuming and prone to failure due to network distortion. A CNN model with a certain parameter range is trained on top of the initial prediction resource pool to act as a proxy for the finite element model and complete the simulation prediction process. This step includes three aspects: network construction, parameter optimization, and overfitting control. Specifically, different numbers of network layers are used to learn the data distribution until the network reaches a complexity that meets the data requirements (overfitting). Overfitting is controlled by sample augmentation and adjusting dropout parameters to improve the model's generalization ability.

[0046] Regarding step 108: It is understandable that when the performance results of the new representative volumetric unit sample are compared with the target performance, the fiber aspect ratio is adjusted using the formula shown in step 102 and input into the reinforcement learning model for a new round of iteration until the microstructure of the composite material can automatically converge to the given range of mechanical properties of the fiber and matrix, thus obtaining the output of the representative volumetric unit model that meets the target performance.

[0047] In some implementations, to accelerate model training and adapt the model to a larger parameter space, a design transfer learning module is incorporated. Through feature transfer, network adjustments are achieved during training with few samples. For example, in reinforcement learning models like DQN, when the aspect ratio of the model fibers changes, the fiber intersection rules are reset, the parameters of the DQN convolutional layers are kept unchanged, and the parameters of the fully connected layers are trained until convergence. In prediction models, the parameters of the prediction model's convolutional layers are kept unchanged, the parameters of the fully connected layers are trained until the model overfits, and the dropout parameters are adjusted until the model achieves maximum accuracy. Throughout this process, new data generated in each iteration can be added to the sample library; an continuously expanding sample library reduces design costs.

[0048] In this embodiment, the effectiveness of the proposed solution was verified. Representative volume element model parameters are defined as follows: RVE side length L1=L2=60mm. Fiber width is 1.5mm, number N=15. Fiber elastic modulus, Poisson's ratio, and matrix elastic modulus, Poisson's ratio are as follows: Ef ~ U(200,350), Em ~ U(35,55), νf ~ U(0.1,0.25), νm ~ U(0.25,0.35). Here, Ef and Em represent the elastic modulus of the fiber phase and matrix phase, respectively, and νf and νm represent the Poisson's ratio of the fiber phase and matrix phase, respectively. The elastic modulus and Poisson's ratio of the fiber and matrix phases are randomly selected within a given interval, and the probability of any value occurring within that interval is equal. The fiber aspect ratio λ∈[8,14] is randomly distributed in a two-dimensional plane. Fiber motion is discretized into 25 central positions and six orientations evenly distributed in [0,π]. Studies have shown that fiber position has little impact on the accuracy of performance prediction. The following sections use model material parameters (Ef, Em) and geometric parameters (fiber aspect ratio λ, fiber angle) as design variables.

[0049] Suppose the target properties of the random short fiber composite material are elastic moduli in two directions: E1 = 40 GPa and E2 = 42 GPa. The results are given using both material parameter iteration strategies and geometric parameter iteration strategies, as follows: Figure 4 The resulting image is shown.

[0050] Run the program and set the precision threshold to 2%. Figure 4Iterative process graphs for two strategies are presented. The horizontal axis represents the number of iterations, and the vertical axis represents the error loss relative to the given performance, defined as follows. It can be seen that the loss gradually decreases with increasing iteration count. After approximately 15 iterations, the iteration stops, and representative volumetric element model images of each phase material property obtained through iteration and fiber geometry parameter iteration, respectively, are obtained as shown below. Figure 5 and Figure 6 As shown, due to the different parameter adjustments in two segments during the iteration process, two different trends will appear before and after the node with loss=5%. Compared with the fiber geometry parameter iteration strategy, the material parameter iteration curve is smoother. This is because the performance of the RSFC model has a linear relationship with the material parameters of each phase, but a nonlinear relationship with parameters such as fiber aspect ratio and orientation.

[0051] To verify the reliability of the results, the geometric model of the fiber distribution results was reconstructed using finite element software, and the corresponding effective performance was simulated. Figure 4 The dashed lines in the diagram show that the error between the results obtained from this scheme and the simulation results is within 5%. Compared with the geometric model of fiber distribution reconstructed by finite element software, this scheme has higher simulation accuracy and takes less time.

[0052] Please refer to Figure 7 This invention provides an intelligent design device for random short fiber composite material structures, used to implement the steps of any method embodiment in the specification. The device includes: Define module 701 to define the fiber aspect ratio range in the reinforcement learning model used for modeling representative volumetric units, and to define the mechanical property range of the fiber and matrix in the prediction model used for performance prediction of representative volumetric units. The adjustment module 702 is used to set the target performance of the random short fiber composite material. It searches for the representative volumetric unit sample that is closest to the target performance in the prediction experience pool, compares the representative volumetric unit sample with the target performance, and adjusts the fiber aspect ratio input to the reinforcement learning model. The initial batch of samples in the prediction experience pool is obtained by combining the random adsorption method and the finite element method. Each sample is stored as a quaternion, which consists of the image name of the representative volumetric unit and the elastic modulus performance. Modeling module 703, the reinforcement learning model is based on a predefined Markov quadruple, calculates the action value mapping function value when each fiber is filled at different positions and angles, selects the position and angle corresponding to the maximum action value mapping function to fill the current fiber, and obtains a new representative volume unit sample; Prediction module 704 is used to adjust the mechanical properties of the fiber and matrix of the new representative volumetric unit sample using a prediction model, so as to obtain the performance results of the new representative volumetric unit sample. The iteration module 705 compares the performance results of the new representative volumetric unit samples with the target performance to adjust the fiber aspect ratio input to the reinforcement learning model and perform a new round of iteration until a representative volumetric unit model output that meets the target performance is obtained.

[0053] It should be noted that the intelligent design device for random short fiber composite material structures provided in the above embodiments is only an example of the division of the above functional units. In practical applications, the above functions can be assigned to different functional units as needed, that is, the internal structure of the device can be divided into different functional units to complete all or part of the functions described above. In addition, the above device embodiments and method embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0054] Embodiments of this application also provide a computer device, please refer to... Figure 8 The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, at least one program, code set or instruction set being loaded and executed by the processor to implement the intelligent design method for random short fiber composite material structures provided in the above-described method embodiments.

[0055] The embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the intelligent design method for random short fiber composite material structures provided in the above-described method embodiments.

[0056] Embodiments of this application also provide a computer program product, which includes a computer program. A processor of a computer device reads the computer program from a computer-readable storage medium and executes the computer program, causing the computer device to perform any of the random short fiber composite material structure intelligent design methods described in the above embodiments.

[0057] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0058] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of the embodiments of this application.

[0059] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0060] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for intelligent design of random short fiber composite material structures, characterized in that, The method includes: In the reinforcement learning model used for modeling representative volumetric units, the fiber aspect ratio range is defined; in the prediction model used for performance prediction of representative volumetric units, the mechanical property range of the fiber and the matrix is ​​defined. A target performance of a random short fiber composite material is set, and a representative volumetric unit sample that is closest to the target performance is found in the prediction resource pool. The representative volumetric unit sample is compared with the target performance to adjust the fiber aspect ratio and input it into the reinforcement learning model. The initial batch of samples in the prediction resource pool is obtained by combining the random adsorption method and the finite element method. Each sample is stored as a quaternion, which consists of the image name of the representative volumetric unit and the elastic modulus performance. The reinforcement learning model is based on a predefined Markov quadruple and calculates the action value mapping function value when each fiber is filled at different positions and angles. The position and angle corresponding to the maximum action value mapping function are selected to fill the current fiber, thus obtaining a new representative volume unit sample. The mechanical properties of the fiber and matrix of the new representative volumetric unit sample are adjusted using the prediction model to obtain the performance results of the new representative volumetric unit sample. The performance results of the new representative volumetric unit samples are compared with the target performance to adjust the fiber aspect ratio input to the reinforcement learning model for a new round of iteration until a representative volumetric unit model output that meets the target performance is obtained.

2. The method as described in claim 1, characterized in that, The representative volumetric unit samples are compared with the target performance to adjust the fiber aspect ratio input to the reinforcement learning model, including: In the formula, The aspect ratio of the fiber. and The performance results for representative volumetric unit samples are the elastic modulus in the first direction and the elastic modulus in the second direction, respectively. and To represent the target performance, the target elastic modulus is defined as the target elastic modulus in the first direction and the target elastic modulus in the second direction. For fiber angle.

3. The method as described in claim 1, characterized in that, The reinforcement learning model is obtained by training a pre-trained convolutional neural network on the constructed reinforcement learning network. The convolutional neural network is used to determine whether there are fiber crossings in the input fiber distribution image in order to calculate the reward value.

4. The method as described in claim 3, characterized in that, The reinforcement learning model is trained in the following manner: Parameters are initialized for the reinforcement learning network and the convolutional neural network; Generate an initial state sample with 1 fiber. For each initial state sample, perform the following: Using a predefined action space, fill the current initial state sample with the next fiber at different positions and angles to generate several state samples with a fiber count of 2; All state samples are input into the convolutional neural network and the reinforcement learning network respectively; The convolutional neural network extracts features from the state samples and determines whether there are fiber crossings in the input state samples. The reward values ​​of each state sample are then input into the reinforcement learning network. The reinforcement learning network calculates the action value mapping function value based on each state sample, calculates the corresponding action value true value using the reward value of each state sample input by the convolutional neural network, calculates the loss function using the action value true value and the action value mapping function value, and adjusts the network parameters of the reinforcement learning network. Based on probability coefficients, select from several state samples with 2 fibers and add them to the experience pool until all initial state samples are filled and trained. For each state sample with 2 fibers in the experience pool, the next fiber at different positions and angles is filled into the state sample using a predefined action space to generate several state samples with 3 fibers. Then, the process jumps to input all state samples into the convolutional neural network and the reinforcement learning network respectively until the set number of fibers is reached, thus completing this round of training of the reinforcement learning network.

5. The method as described in claim 4, characterized in that, When a state sample represents a three-dimensional representative volume element model, the pixel grayscale values ​​in the state sample characterize the orientation information of the three-dimensional representative volume element model. The correspondence between the pixel grayscale values ​​in the state sample and the orientation information is as follows: In the formula, The midpoint of the state sample pixel grayscale values, This represents the Z-coordinate value of the point corresponding to the representative three-dimensional volume element model. The directional side length of the representative three-dimensional volume element model. Here is the Z-coordinate value of the first endpoint of the fiber in the three-dimensional representative volume element model. The midpoint of the state sample The distance to the first end of the fiber. Euler angles; The determination of whether fiber crossings exist in the input state sample includes: Based on the grayscale value of each pixel in each fiber in the state sample, calculate the position of each pixel in the three-dimensional representative volume unit model. Calculate whether the distance between the centerlines of any two fibers is greater than the fiber width; If so, confirm that the two fibers do not cross each other; If not, it is determined that the two fibers intersect.

6. The method according to claim 3, characterized in that, The action space is represented by the following formula: In the formula, For the action space, This represents the filling position of the fiber center point in the state sample. Let the angle of the fiber in the state sample be denoted as . The side length of the representative three-dimensional volume element model; The reward value for each state sample is calculated using the following formula: In the formula, For filling action Formed state samples The corresponding reward value, This indicates that the fibers are effective, meaning there is no cross-linking between the fibers. This indicates that the fiber is invalid, meaning that the fiber crosses. This indicates that the fiber filling has reached the required number of fibers, and the modeling process is complete. v>>t>0, u<0.

7. A random short fiber composite material structure intelligent design device, used to implement the steps of the method described in any one of claims 1-6, characterized in that, The device includes: Define the fiber aspect ratio range in the reinforcement learning model used for modeling representative volumetric elements, and define the mechanical property range of the fiber and matrix in the prediction model used for performance prediction of representative volumetric elements. The adjustment module is used to set the target performance of the random short fiber composite material, find the representative volumetric unit sample that is closest to the target performance in the prediction experience pool, compare the representative volumetric unit sample with the target performance, and adjust the fiber aspect ratio input to the reinforcement learning model. The initial batch of samples in the prediction experience pool is obtained by combining the random adsorption method and the finite element method. Each sample is stored as a quaternion, which consists of the image name of the representative volumetric unit and the elastic modulus performance. The modeling module, based on a predefined Markov quadruple, calculates the action value mapping function value for each fiber filled at different positions and angles, and selects the position and angle corresponding to the maximum action value mapping function to fill the current fiber, thereby obtaining a new representative volume unit sample. The prediction module is used to adjust the mechanical properties of the fiber and matrix of the new representative volumetric unit sample using the prediction model, so as to obtain the performance results of the new representative volumetric unit sample. The iteration module compares the performance results of the new representative volumetric unit samples with the target performance to adjust the fiber aspect ratio input to the reinforcement learning model and perform a new round of iteration until a representative volumetric unit model output that meets the target performance is obtained.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1-6.