A Closed-Loop Generative Inverse Design Method for Lattice Structures

Abstract

This application discloses a closed-loop generative inverse design method for lattice structures, relating to the field of structural design technology. The method first establishes a digital topology encoding vector to describe the topology of the lattice structure, and constructs a mapping dataset between the topological features of the lattice structure and mechanical performance indicators based on parametric simulation. Then, it constructs a mechanical performance surrogate model and a conditional variational autoencoder model constrained by target performance conditions, mapping the high-dimensional discrete topological space to a continuous latent feature manifold. Further, it constructs a reinforcement learning agent, using the target performance conditions as the state space input and the continuous latent feature manifold as the action space. The agent predicts performance results and constructs reward feedback based on performance deviations to update the reinforcement learning policy. Finally, it outputs multiple lattice structure schemes that satisfy the target performance conditions but have different topologies. This invention can improve the efficiency and performance matching accuracy of inverse design for complex lattice structures.

Description

Technical Field

[0001] This invention relates to the field of structural design technology, and more specifically to a closed-loop generative inverse design method, apparatus, and equipment for lattice structures. Background Technology

[0002] As a typical type of mechanical metamaterial, lattice structures are widely used in aerospace, transportation, impact protection, and advanced manufacturing due to their lightweight, high specific strength, high specific stiffness, and excellent energy absorption capabilities. With the development of additive manufacturing technology, the processing capability of complex lattice topologies has been continuously improved. How to achieve rapid reverse design of lattice structures according to the target mechanical performance requirements has gradually become an important research direction in the field of structural optimization.

[0003] Existing lattice structure design methods mostly employ parameter traversal, empirical adjustment, or heuristic optimization. These methods involve continuously modifying the structural topology and analyzing the mechanical response using finite element simulation to select structural schemes that meet performance requirements. Such methods typically rely on extensive repetitive calculations, resulting in high computational costs, low search efficiency, and difficulty in quickly converging to the target performance in high-dimensional discrete design spaces.

[0004] In recent years, deep learning-based generative design methods have been increasingly applied to the inverse design of lattice structures. By learning the mapping relationship between structural topology and mechanical properties, structural generation and performance prediction can be achieved. However, most existing generative design methods adopt open-loop generation mechanisms, and their generation process usually lacks real-time physical feedback and dynamic correction capabilities oriented towards target performance, making it difficult for the generated results to accurately meet the preset performance requirements. At the same time, for the same performance target, existing methods are unable to stably output multiple structural schemes with significant topological differences but similar performance, limiting the flexibility of structural design and its adaptability to engineering applications.

[0005] Therefore, there is an urgent need for a closed-loop generative inverse design method for lattice structures that can achieve target performance orientation, has real-time physical feedback capability, and takes into account structural diversity. Summary of the Invention

[0006] The purpose of this invention is to provide a closed-loop generative inverse design method for lattice structures. By constructing a mapping relationship between the topological features of lattice structures and mechanical performance indicators, and by coupling a conditional variational autoencoder model, a mechanical performance surrogate model, and a reinforcement learning agent in a closed loop, the automatic inverse generation of lattice structure schemes from target performance conditions is achieved. Compared with traditional design methods that rely on parameter traversal or empirical adjustment, this application maps a high-dimensional discrete topological space to a continuous latent feature manifold, and uses this continuous latent feature manifold as the action space of the reinforcement learning agent, effectively improving the search efficiency and policy convergence stability of complex lattice structures in the high-dimensional design space. Simultaneously, by providing real-time mechanical performance feedback through the surrogate model and dynamically updating the strategy based on the deviation between the predicted performance results and the target performance conditions, the generated lattice structures can more accurately meet the target mechanical performance requirements, and multiple lattice structure schemes with the same performance constraints but different topological structures are output, improving the automation and design flexibility of lattice structure inverse design.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a closed-loop generative inverse design method for lattice structures, the method comprising: Obtain the target performance conditions of the lattice structure and establish a digital topology encoding vector to describe the topology of the lattice structure; A parameterized simulation process for lattice structures is constructed based on digital topology encoding vectors. Mechanical simulations are performed on multiple topology encoding samples to extract the corresponding mechanical performance indicators and to construct a mapping dataset between the topological features of lattice structures and their mechanical performance indicators. A mechanical performance proxy model is constructed based on the mapping dataset to achieve rapid prediction of the mechanical performance of lattice structures; A conditional variational autoencoder model constrained by target performance conditions is constructed, which maps lattice structure features to continuous latent feature manifolds, and reconstructs latent vectors into lattice structure features through a decoder; Construct a reinforcement learning agent, take the target performance conditions as the input to the state space, take the continuous latent feature manifold as the action space, and use the reinforcement learning agent to output the latent action vector. The latent action vector is input into the conditional variational autoencoder model for decoding to obtain the reconstructed lattice structure features. The mechanical performance of the reconstructed lattice structure features is then predicted using a mechanical performance surrogate model to obtain the predicted performance results. Reward feedback is constructed based on the deviation between the predicted performance results and the target performance conditions to drive the reinforcement learning agent to update the policy until the preset convergence conditions are met. A lattice structure scheme based on the output of the converged reinforcement learning agent that satisfies the target performance conditions.

[0008] Secondly, the present invention also provides a lattice structure closed-loop generative inverse design device, the device comprising: The vector creation module is used to obtain the target performance conditions of the lattice structure and to create a digital topology encoding vector to describe the topology of the lattice structure. The mechanical simulation module is used to construct a parameterized simulation process for lattice structures based on digital topology encoding vectors, perform mechanical simulations on multiple topology encoding samples, extract the corresponding mechanical performance indicators, and construct a mapping dataset between the topological features of lattice structures and mechanical performance indicators. The model building module is used to build mechanical performance proxy models based on the mapping dataset to achieve rapid prediction of the mechanical performance of lattice structures. The feature mapping module is used to construct a conditional variational autoencoder model constrained by the target performance conditions, map the lattice structure features to a continuous latent feature manifold, and reconstruct the latent vectors to lattice structure features through the decoder; The reinforcement learning module is used to construct reinforcement learning agents. It takes the target performance conditions as input to the state space, the continuous latent feature manifold as the action space, and uses the reinforcement learning agent to output latent action vectors. The result prediction module is used to decode the latent action vector input conditional variational autoencoder model to obtain the reconstructed lattice structure features, and to predict the mechanical properties of the reconstructed lattice structure features through a mechanical performance surrogate model to obtain the prediction performance results. The policy update module is used to construct reward feedback based on the deviation between the predicted performance results and the target performance conditions, so as to drive the reinforcement learning agent to update the policy until the preset convergence conditions are met. The scheme output module is used to output a lattice structure scheme that meets the target performance conditions based on the converged reinforcement learning agent.

[0009] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the lattice structure closed-loop generative inverse design method provided in the first aspect.

[0010] The beneficial effects of this invention are as follows: This application provides a closed-loop generative inverse design method for lattice structures. By constructing a mapping relationship between the topological features of the lattice structure and its mechanical performance indicators, and by coupling a conditional variational autoencoder model, a mechanical performance surrogate model, and a reinforcement learning agent in a closed loop, it achieves automatic inverse generation from target performance conditions to lattice structure schemes. Compared to traditional design methods that rely on parameter traversal or empirical adjustment, this application maps a high-dimensional discrete topological space to a continuous latent feature manifold, and uses this continuous latent feature manifold as the action space of the reinforcement learning agent, effectively improving the search efficiency and policy convergence stability of complex lattice structures in the high-dimensional design space. Simultaneously, by providing real-time mechanical performance feedback through the surrogate model and dynamically updating the strategy based on the deviation between the predicted performance results and the target performance conditions, the generated lattice structure can more accurately meet the target mechanical performance requirements, and outputs multiple lattice structure schemes that satisfy the same performance constraints but have different topological structures, thereby improving the automation level and design flexibility of lattice structure inverse design.

[0011] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a closed-loop generative inverse design method for lattice structures according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a topology encoding and decoding framework for an axially extended lattice structure according to an embodiment of the present invention; Figure 3 This is a compression model diagram of an axially extended lattice structure according to an embodiment of the present invention; Figure 4 This is a diagram showing the typical force-displacement response, displacement contour lines, and performance indicators of an axially extended lattice structure under compression at different deformation stages, as illustrated in an embodiment of the present invention. Figure 5 This is a schematic diagram of the system architecture of a deep residual proxy model for lattice attribute prediction according to an embodiment of the present invention; Figure 6 This is a schematic diagram of a PC-VAE training framework according to an embodiment of the present invention; Figure 7 This is a schematic diagram of a reinforcement learning framework based on PPO structure inverse design, as shown in an embodiment of the present invention. Figure 8 This is a schematic diagram of the generative inverse design results for achieving precise single-objective optimization with a peak stress of 4750 MPa, as shown in an embodiment of the present invention. Figure 9 This is a schematic diagram of the single-objective continuous interval control generative inverse design result with energy absorption between 4500-5000 mJ, as shown in an embodiment of the present invention. Figure 10 This is a schematic diagram of the generative inverse design results for precise multi-objective control with a peak stress of 4500 MPa and an energy absorption of 7000 mJ, as shown in an embodiment of the present invention. Figure 11 This is a schematic diagram of a closed-loop generative inverse design device for a lattice structure according to an embodiment of the present invention. Figure 12 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0013] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics; however, not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0015] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0016] In some embodiments, a closed-loop generative inverse design method for lattice structures is provided, the overall execution flow of which is as follows: Figure 1 As shown, this method aims to automatically generate lattice structure schemes that meet performance requirements and possess topological diversity based on user-given target performance conditions. Here, "lattice structure" refers to a lightweight periodic structure formed by multiple nodes and connecting rods according to a preset topological relationship, which typically possesses excellent specific strength, buffer energy absorption, and impact resistance. This embodiment uses an axially extended lattice structure as a specific example for illustration, but those skilled in the art should understand that this method is also applicable to other metamaterial structures with discrete topological expression characteristics.

[0017] S101, obtain the target performance conditions of the lattice structure, and establish a digital topology encoding vector to describe the topology of the lattice structure.

[0018] Among them, "target performance conditions" refer to the constraints proposed by the user for the final mechanical performance of the lattice structure, which are used as the optimization direction for the subsequent inverse design process. They may include one or more of the following: peak stress, buffer energy absorption, stiffness, and buckling threshold.

[0019] To enable reinforcement learning agents to handle lattice structures with complex connectivity, the original geometric structure must first be converted into a computer-processable digital representation. Here, "digital topological encoding vector" refers to the feature vector formed by discretely representing the node connection logic, path states, and spatial configuration in the lattice structure. Its function is to transform the complex three-dimensional structure, which is originally difficult to process directly, into a unified digital input format, facilitating subsequent parametric modeling, surrogate model training, and reinforcement learning policy search.

[0020] In this embodiment, reference Figure 2 , Figure 2 This diagram illustrates a topology encoding and decoding framework for axially extended lattice structures. Specifically, a topology encoding and decoding framework is established for axially extended lattice structures. Multiple independent topological paths in the lattice structure are each set as a path control subset. Each path control subset includes connection mode variables and node selection variables. The connection mode variables describe whether adjacent nodes are connected and the connection method, such as the linear connection state within a dual-path system. The node selection variables describe the spatial extension relationship of the structure among candidate nodes, such as the member connection direction, node position offset, and path selection logic. By combining and encoding these variables, a digital topology encoding vector describing the overall structural topology can be formed.

[0021] For example, the axially extended lattice structure is formed by linearly stretching a two-dimensional cross-section along the axial direction. To establish a unified geometric reference system, the four corner points of the two-dimensional cross-section are first defined as fixed boundary nodes and sequentially numbered 1, 2, 3, and 4, connected to form a closed outer frame to provide basic structural stability. Inside the outer frame, a structure consisting of nine candidate nodes is arranged. The regular grid has nodes numbered sequentially from 5 to 13, serving as discrete control points to control topology changes.

[0022] The internal topology of the lattice structure is constructed from two spatially independent paths: the first path connects node 1 and node 3, and the second path connects node 2 and node 4. To transform this topological construction into a computer-recognizable digital representation, this embodiment introduces a 10-dimensional discrete encoding vector. The vector is symmetrically partitioned into two subsets, each containing 5 variables: the first 5 variables... Corresponding to the first path, the last 5 variables This corresponds to the second path. Among them, This is a connection mode variable, i.e. a geometric logic switch. When the value is 1, it indicates that the connection between two adjacent nodes is a straight line, and when the value is 0, it indicates that the connection is a smooth curve. Variables, or node indices, are selected for nodes following a row-by-row progression rule to ensure structural continuity. Their specific value range is constrained by the number of rows in the candidate grid they occupy: First row variables... The second row of variables The third row of variables .

[0023] Based on the above encoding rules, complex topological arrangements can be precisely converted into unique digital codes. Figure 2 Taking the demonstrated encoding / decoding framework as an example, its corresponding topological coding vector is: The specific decoding mapping logic is as follows: For the first path, its control variable is... The node index variable indicates that the path starts from fixed node 1, passes through candidate node 6 in the first row, candidate node 10 in the second row, and finally reaches fixed node 3. Connection pattern variable. This indicates that all connections between nodes 1 and 6, 6 and 10, and 10 and 3 use straight line segments; for the second path, its control variable is... The path starts from fixed node 2, sequentially passes through candidate node 6 in the first row and candidate node 8 in the second row, and finally reaches fixed node 4; connects pattern variables. This indicates that the first part uses a smooth curve for connection, while the second part uses a straight line for connection.

[0024] Two paths are generated independently and interweave, ultimately forming a unified two-dimensional topological section. This section is then stretched axially to parametrically generate a three-dimensional lattice unit cell geometric model. Through this encoding mechanism, each path has three choices from three rows of candidate nodes, for a total of... There are 4 possible combinations, and the two connecting segments have 4 possible connecting line configurations, so a single path can evolve into 100 different combinations. There are several independent topologies. Since the two paths are independent, the total number of theoretical configurations contained in the complete discrete design space constructed in this embodiment is precisely [number missing]. kind.

[0025] The beneficial effect of the above process is that by uniformly transforming complex lattice structures into digital topological encoding vectors, the transformation from discrete topological logic to a continuous computable expression is realized, enabling different topological configurations to be described in a unified mathematical space. This provides a foundation for establishing the mapping relationship between lattice structures and mechanical properties, while avoiding the inefficiency caused by repeatedly adjusting structural parameters in traditional manual modeling methods.

[0026] S102, based on the digital topology encoding vector, constructs a parameterized simulation process for lattice structures, performs mechanical simulations on multiple topology encoding samples, extracts the corresponding mechanical performance indicators, and constructs a mapping dataset between the topological features of the lattice structure and the mechanical performance indicators.

[0027] The "parametric simulation process" refers to an integrated execution mechanism that can automatically complete geometric modeling, finite element solution, and performance extraction based on input codes. Its purpose is to automatically generate high-fidelity training data to reduce the cost of repetitive manual modeling. In this embodiment, reference is made to... Figure 3 as well as Figure 4 , Figure 3 This is a diagram of a compression model of an axially extended lattice structure. Figure 4 To illustrate the typical force-displacement response, displacement contour lines, and performance indicators of axially extended lattice structures under different compression deformation stages, an automated parametric modeling process is established based on digital topological coding vectors. When different codes are input, the system automatically triggers parametric modeling scripts to generate the corresponding 3D model of the axially extended lattice structure.

[0028] Subsequently, hierarchical sampling is performed on the digitized topology encoding vector to generate multiple topology encoding samples. Here, "hierarchical sampling" refers to extracting representative structural samples from the overall design space while ensuring balanced coverage of various topological features, thereby improving the diversity and generalization ability of the training data. Specifically, this embodiment constructs multiple topology encoding samples by combining sampling of different path combination patterns, node spatial positions, and geometric parameter variations, and calls the finite element simulation process one by one.

[0029] In the finite element analysis phase, an axial compressive load is applied to the generated three-dimensional lattice structure, and the response of the structure at different compressive displacement stages is obtained through the solver. (Reference) Figure 4 This allows us to obtain force-displacement response curves and displacement contour distributions at different deformation stages. Then, mechanical performance indicators are extracted based on the compression response curves. Here, "mechanical performance indicators" are used to characterize the actual mechanical behavior of the lattice structure. In this embodiment, at least peak stress and energy absorption indicators are included. Peak stress reflects the structure's load-bearing limit; energy absorption measures the structure's energy dissipation capacity during compression.

[0030] Specifically, the 10-dimensional discrete topological coding space theoretically contains 11,664 different topological configurations. Considering computational cost and global representation requirements, the sampling size is set to... Dimension Using the Latin hypercube sampling algorithm, the first... The sample at the th Continuous sampled values ​​on dimension Calculate using the following formula: ; In the formula, This represents a random permutation of integers between 1 and N. Indicates in Random numbers that follow a uniform distribution within the interval. After sampling, the continuous variable... The data is rounded to the nearest discrete integer level. If duplicate topological codes are generated during the discretization mapping, an iterative resampling process is triggered to remove and complete them until 2100 completely unique and independent structurally encoded sample sequences are obtained. The generated dataset exhibits a very high uniform distribution of design variables at the discrete level, and the correlation coefficients between variables are close to 0, ensuring the independence of global feature sampling.

[0031] An automated control script based on Python was developed to construct a quasi-static uniaxial compression simulation solver for lattice structures. The script automatically reads the 2100 sets of discrete codes for the lattice structures generated by the aforementioned layered sampling and repeatedly calls the Abaqus simulation kernel in the background. In each round of the solution, the script automatically generates the corresponding mesh model based on the current code and applies a downward displacement load of 4 mm, outputting the mechanical test response of the lattice structure. After the solution is completed, the script automatically captures and extracts the reaction force-displacement response curves of each lattice structure throughout the compression process, identifies the maximum stress value, and names it the peak stress. Simultaneously, the force-displacement curve is integrated up to the point before densification to calculate the strain energy absorption, which is termed buffer energy absorption. .

[0032] The automated data integration program performs bidirectional lateral pairing of 2100 sets of input topological codes with their corresponding output mechanical properties. The paired samples are then compiled into a standard data matrix and stored uniformly in .csv format.

[0033] The beneficial effect of the above process is that by constructing an automated parametric simulation process, the automatic mapping and acquisition of lattice structure data from topology encoding to mechanical performance data can be realized. This can effectively reduce the time cost of a large number of repetitive finite element modeling and simulation in traditional manual trial-and-error design, and improve the accuracy and generalization ability of subsequent model training by covering datasets with multiple topology combination modes.

[0034] S103, a mechanical performance proxy model is constructed based on the mapping dataset to achieve rapid prediction of the mechanical performance of lattice structures.

[0035] Among them, the "mechanical performance proxy model" refers to a prediction model used to replace the traditional finite element solution process. Its core function is to achieve millisecond-level performance prediction by learning the mapping law between structural features and mechanical response in historical samples, thereby avoiding the frequent use of computationally expensive finite element analysis in the reinforcement learning process.

[0036] In this embodiment, reference Figure 5 , Figure 5 This diagram illustrates the system architecture of a deep residual surrogate model for lattice attribute prediction. Before constructing the surrogate model, physical feature engineering is performed on the digitized topology encoding vector to obtain continuous physical feature vectors. Here, "physical feature engineering" refers to further converting discrete topological information into continuous physical feature representations more suitable for neural network learning. For example, node spatial locations, link connectivity, and geometric parameters can be uniformly mapped into continuous feature vectors containing both topological logical features and geometric attribute features, thereby enhancing the model's ability to learn structure-performance relationships.

[0037] Subsequently, the continuous physical feature vectors are input into a deep residual neural network for training. Here, "deep residual neural network" refers to a neural network model that introduces a residual connection mechanism into a deep network structure. Its role is to reduce the problems of gradient vanishing and training degradation during the training process of deep networks, thereby enhancing the learning stability of complex structural mechanical mapping relationships. In this embodiment, the surrogate model extracts the coupling features between the lattice structure and mechanical properties layer by layer through multiple residual modules, and uses the peak stress and buffer energy absorption index in the mapping dataset as supervision signals to iteratively update the network parameters.

[0038] After meeting the preset performance evaluation metrics, the trained network parameters are frozen. "Freezing" here means fixing the network weight parameters so that they are used only as predictors in subsequent closed-loop reinforcement learning processes and no longer participate in parameter updates, thus ensuring the stability of the reinforcement learning reward feedback source. For example, when a structure to be predicted is input into the surrogate model, the predicted peak stress and predicted buffer energy absorption can be quickly output to replace the originally time-consuming finite element calculations.

[0039] For example, directly using a 10-dimensional discrete topological encoding vector as network input can cause deep neural networks to struggle to converge when faced with highly nonlinear mechanical abrupt changes. Therefore, this embodiment develops a physical feature engineering module to explicitly expand and transform the 10-dimensional discrete encoding into a 19-dimensional continuous physical feature vector. Specifically, the 19-dimensional continuous physical feature vector is composed of four dimensions of path connectivity pattern features (extracting the linear switching states within the two paths), twelve dimensions of key node spatial coordinate features (quantitatively calculating the Cartesian coordinates of the six key nodes activated in the candidate grid), and three dimensions of global geometric attribute features (including the total geometric arc length of the two connected paths and the Euclidean distance between their geometric centers). After feature extraction, the 19-dimensional continuous physical feature vector is scaled across the entire set using the Z-score normalization algorithm. By calculating the mean and standard deviation of each feature dimension across the entire set, the original feature values ​​are subtracted from the mean and divided by the standard deviation, forcing the overall mean to zero and the variance to one. This scaled and normalized 19-dimensional continuous physical feature vector then serves as the original input signal for the deep residual neural network.

[0040] like Figure 5 As shown, the input mapping layer first receives a 19-dimensional normalized continuous physical feature vector. This vector is then projected onto a 128-dimensional high-dimensional latent space using a fully connected linear transformation matrix, and a SiLU activation function is introduced for nonlinear transformation. Following the input mapping layer is the core feature extraction module, which consists of four sequentially stacked, structurally identical residual blocks connected in series. Each residual block contains two fully connected layers (each with 128 neurons), a batch normalization layer, and a SiLU activation function. A dropout rate of 0.2 is set between layers to effectively suppress overfitting. Finally, connected to the end of the core feature extraction module is a dimensionality reduction output layer. This layer uses a multilayer perceptron to linearly reduce the 128-dimensional deep high-order features to a 2-dimensional space, simultaneously outputting the predicted buffer absorption value and the predicted peak stress value. These two values ​​are then combined to construct a 2-dimensional prediction performance vector. .

[0041] To ensure that the underlying geometric features are not lost during the training of the deep network, the forward propagation mathematical mapping logic of each residual block in the core feature extraction module is defined as follows: ; In the formula, Indicates input to the first The 128-dimensional hidden feature vector of each residual block. This indicates that the residual block contains a nonlinear residual transformation operator consisting of two fully connected layers, batch normalization, and a discard layer. These are the weights and bias parameters to be learned within the operator. This is the activation function for the outer SiLU layer. This outputs a new 128-dimensional feature vector for the residual block and passes it to the next layer. (The input is then passed to the next layer.) The direct jump summation to the transformation feature constitutes an identity short-circuit mapping, ensuring the smooth flow of gradients during backpropagation.

[0042] Considering that fine-tuning the lattice structure topology may lead to sudden jumps in nonlinear mechanical response, to improve the robustness of the surrogate model to outliers, the conventional mean squared error is abandoned, and mean absolute error is chosen as the training loss function. Used to calculate the predicted performance vector With the finite element real performance labels in the .csv file The residual between them is defined by the formula: ; In the formula, The sample batch size is used for a single training iteration. The hyperparameter configuration of the network is as follows: an adaptive moment estimator with weight decay is used for parameter optimization, and overfitting is suppressed by introducing a weight decay coefficient. The initial learning rate is set to... A cosine annealing scheduler is used to dynamically and smoothly decay the learning rate throughout the training cycle. The batch size for each training session is set. The total number of training iterations is set to 2000.

[0043] The training and deployment process of the agent model is based on Figure 5 The deep residual proxy model system architecture is shown below. First, the program reads a matrix file containing 2100 data sets and randomly divides it proportionally into a training set (1600 samples), a validation set (200 samples), and a test set (300 samples). During training, the standardized physical feature vectors of the training set are input into the network, and multiple rounds of forward computation and parameter updates are performed according to the propagation mapping and loss function. Simultaneously, the validation set is used to monitor prediction errors to prevent overfitting. When training reaches a preset 2000 rounds and meets the preset mechanical accuracy index, the network is considered to have reached full convergence. At this point, the full layer parameters of the current network are exported via a save command, and the weights and bias parameters of all layers are frozen. The resulting mechanical performance proxy model can reduce the time required for mechanical evaluation of a single lattice configuration to an extremely short time, directly serving as a real-time evaluation interface for reinforcement learning environment interaction.

[0044] The beneficial effect of the above process is that by establishing a mechanical performance proxy model, the nonlinear relationship between complex lattice structures and mechanical responses can be quickly approximated using deep networks, realizing the transformation from traditional time-consuming finite element analysis to a fast prediction mode. This significantly reduces the computational overhead in reinforcement learning closed-loop optimization and provides stable data support for subsequent real-time physical feedback.

[0045] S104. Construct a conditional variational autoencoder model constrained by the target performance conditions, map the lattice structure features to a continuous latent feature manifold, and reconstruct the latent vectors to lattice structure features through a decoder.

[0046] Among them, the "Conditional Variational Autoencoder Model" is a generative neural network model that combines feature compression and conditional generation capabilities. Its core function is to compress and map the original high-dimensional, discrete, and discontinuous lattice topological space to a low-dimensional continuous latent space, and to achieve controllable generation of structural features under the constraints of target performance conditions. Here, the "latent feature manifold" refers to the continuous low-dimensional feature space learned by the neural network. In this space, the lattice structures corresponding to adjacent positions have a continuous relationship of change in geometric topology and mechanical properties, thereby providing a stable continuous action search space for reinforcement learning agents.

[0047] In this embodiment, reference Figure 6 , Figure 6 The diagram illustrates the PC-VAE training framework. First, the previously obtained continuous physical feature vectors are concatenated with the target performance conditions. "Feature concatenation" refers to merging multiple input variables into a unified input vector along the feature dimension, enabling the network to simultaneously learn the coupling relationship between structural topology features and target performance.

[0048] Next, the joint input vector is fed into a probabilistic encoder for feature compression. Here, the "probabilistic encoder" refers to an encoding network used to learn the probability distribution features of the latent space, its function being to compress and map the original high-dimensional structural features to a low-dimensional latent space. In this embodiment, the probabilistic encoder adopts a multilayer perceptron structure, which contains multiple fully connected layers and utilizes a nonlinear activation function to extract latent topological patterns from the high-dimensional geometric features. After encoding, the mean vector and log-variance vector in the latent space are output.

[0049] The "mean vector" represents the center position of the current structural sample in the latent space, while the "log-variance vector" characterizes the uncertainty range of the latent space distribution. Since direct random sampling prevents the network from performing gradient backpropagation, this embodiment further employs a reparameterized sampling mechanism to generate the latent vector.

[0050] Specifically, reparameterized sampling is performed based on the mean vector, log-variance vector, and random noise to obtain a latent vector that satisfies the following relationship: z = μ + σ⊙ε; where μ represents the mean vector, σ represents the standard deviation vector determined by the log-variance vector, ε represents the noise vector randomly sampled from the standard normal distribution, and ⊙ represents the element-wise multiplication operation.

[0051] Here, "reparameterized sampling" refers to converting the random sampling process into a differentiable function expression to ensure stable gradient propagation during network training. For example, after the encoder outputs the mean vector of a certain structural sample, random perturbations can be superimposed around this mean to obtain multiple latent vectors with similar topological features but subtle differences. This not only enhances the continuity of the latent space but also allows small changes in latent variables to correspond to the continuous evolution of the lattice structure topology.

[0052] Subsequently, the obtained latent vectors and target performance conditions are input again into the conditional decoder for feature reconstruction. Here, the "conditional decoder" refers to a generative network used to recover high-dimensional structural features from low-dimensional latent vectors; its function is to recover the corresponding lattice structure representation based on the position in the latent space. In this embodiment, the conditional decoder utilizes a multi-layer fully connected network to perform feature upscaling and ultimately outputs the reconstructed physical feature vectors.

[0053] Considering that the output features simultaneously contain discrete topological logic and continuous geometric coordinates, this embodiment further employs a hybrid output activation mechanism. Specifically, for the topological switch variables representing the internal connection state of the two paths, a sigmoid activation function is used to constrain them, ensuring that the output value is between 0 and 1, thereby satisfying the topological connectivity logic requirements; for the continuous variables representing the spatial coordinates and geometric dimensions of key nodes, a linear mapping method is used for output to avoid geometric information distortion caused by nonlinear truncation.

[0054] For example, when a latent vector corresponds to a high-energy-absorbing structural region, the decoder can automatically generate a lattice topology with complex path connectivity and a large deformation space. As the latent vector gradually moves towards a high-stiffness region, it gradually generates a structural configuration with denser member distribution and more stable node support. Thus, it can be seen that in the latent feature manifold, different regions can correspond to structural topological clusters guided by different mechanical properties.

[0055] To further ensure that the generated structure meets the real physical laws, this embodiment also constructs a composite loss function coupled with physical consistency verification. The "composite loss function" refers to a joint optimization objective that simultaneously considers geometric reconstruction error, the continuity of potential spatial distribution, and mechanical performance deviation. Its role is to guide the conditional variational autoencoder model to meet actual mechanical constraints while ensuring structural generability.

[0056] Specifically, the composite loss function is: ; in, For geometric reconstruction loss term, Here, β is the KL divergence term, and β is the balance weight coefficient. For the physical consistency constraint loss term, As a surrogate model for mechanical properties, The feature vector reconstructed by the decoder. Let be the given condition vector; λ be the penalty weight coefficient. The geometric reconstruction loss term constrains the geometric deviation between the decoded structure and the original structure; the distribution regularization divergence term constrains the latent space distribution to approximate a standard normal distribution, enhancing the continuity of the latent space; the physical consistency constraint loss term, by calling the aforementioned mechanical performance proxy model with frozen parameters, performs real-time performance prediction on the decoded structure and calculates the deviation between the predicted performance result and the target performance condition, ensuring that the generated structure conforms to real mechanical laws.

[0057] For example, when the structure generated by the decoder can geometrically reconstruct the original topology, but its predicted peak stress deviates too much from the target value, the physical consistency constraint loss term will increase the corresponding penalty, thereby forcing the network to readjust the potential spatial mapping relationship in order to gradually learn the structure generation rules that conform to the target performance constraints.

[0058] Subsequently, the conditional variational autoencoder model was trained iteratively in multiple rounds using the aforementioned mapping dataset, and the network parameters of the probabilistic encoder and conditional decoder were saved after the composite loss function converged to form a continuous latent feature manifold.

[0059] In this embodiment, a total of 2100 sets of high-fidelity finite element samples were used to train the model. In the early stage of training, there may be a competitive relationship between geometric reconstruction error and physical constraint error. That is, although the network can reconstruct the geometric structure well, there are still deviations in the prediction of mechanical performance. As the training progresses, the network will gradually learn the implicit correlation between structural topology and performance, so that the composite loss function will decrease steadily and eventually converge.

[0060] The beneficial effect of the above process is that, by constructing a conditional variational autoencoder model constrained by target performance conditions, a mapping transformation from a high-dimensional discrete topological space to a low-dimensional continuous latent space is achieved, enabling complex lattice structures to be smoothly searched and controllably generated in the continuous space. Simultaneously, by introducing a mechanical performance surrogate model to participate in physical consistency constraints, the latent space not only possesses geometric continuity but also the ability to be constrained by real mechanical laws, thus providing a foundation for the stable optimization of reinforcement learning agents in the continuous action space.

[0061] S105, construct a reinforcement learning agent, take the target performance conditions as the input to the state space, take the continuous latent feature manifold as the action space, and use the reinforcement learning agent to output the latent action vector.

[0062] Here, "reinforcement learning agent" refers to an autonomous optimization model that can continuously adjust its strategy through environmental feedback. Its role is to automatically search for potential structural solutions that meet the requirements based on the target performance conditions. "State space" is used to describe the current optimization objective, while "action space" is used to describe the structural search actions that the agent can perform. In this embodiment, the lattice structure inverse design process is defined as a Markov decision process in a continuous state space.

[0063] Specifically, the target performance conditions are constructed as the state space input. Then, the continuous latent feature manifold learned by the aforementioned conditional variational autoencoder model is defined as the action space of the reinforcement learning agent. Here, "action space" refers to the continuous variable space in which the agent can perform search and decision-making. In this embodiment, the actions output by the reinforcement learning agent are the continuous latent vectors in the latent space.

[0064] Compared to the traditional method of directly searching for structural schemes in the discrete topological coding space, this embodiment does not directly output the specific link connection states. Instead, it first searches for latent action vectors in a continuous latent manifold, and then recovers the specific lattice structure through a decoder. Since adjacent positions in the latent manifold correspond to structures with continuously changing topology, it avoids the large-scale search discontinuity problem that occurs in the traditional discrete space.

[0065] For example, when a reinforcement learning agent makes a small adjustment to an action vector in the latent space, the resulting lattice structure usually only shows changes in local rod thickness, node offset, or gradual changes in connection patterns, without any completely unrelated topological jumps. This can significantly improve gradient stability and search convergence ability during the reinforcement learning policy optimization process.

[0066] The beneficial effect of the above process is that by using the continuous latent feature manifold obtained by constructing the conditional variational autoencoder model as the action space of the reinforcement learning agent, the transformation from traditional high-dimensional discrete topology search to continuous low-dimensional latent search is realized. This can effectively reduce the search difficulty of reinforcement learning in complex structure design problems and improve the policy stability and optimization efficiency in the inverse design process of lattice structures.

[0067] S106, the latent action vector is input into the conditional variational autoencoder model for decoding to obtain the reconstructed lattice structure features, and the mechanical performance of the reconstructed lattice structure features is predicted through the mechanical performance surrogate model to obtain the prediction performance results.

[0068] "Decoding" refers to restoring the low-dimensional latent action vectors output by the reinforcement learning agent into a lattice structure representation that can be practically modeled and analyzed. Its role is to establish the correspondence between reinforcement learning actions and actual structural topology. "Predicted performance results" refers to the structural mechanical response results obtained by quickly calculating through the surrogate model, which is used to replace traditional finite element analysis to achieve real-time performance feedback in the closed-loop process of reinforcement learning.

[0069] In this embodiment, reference Figure 7 , Figure 7 This diagram illustrates a reinforcement learning framework based on PPO structure inverse design. After the reinforcement learning agent outputs latent action vectors in a continuous latent feature manifold, it first inputs these latent action vectors along with the target performance conditions into a trained conditional decoder. The conditional decoder automatically reconstructs the corresponding physical feature vectors based on the positions of the latent vectors in the continuous latent space, and further recovers the corresponding digital topological codes.

[0070] Here, "digital topology encoding recovery" refers to converting the continuous features output by the decoder back into a topological representation that can be used for geometric modeling. For example, for topological switch variables in the decoded output, the connection between adjacent nodes can be determined based on their output probability values; for key node coordinate variables, their geometric position parameters can be directly read, thereby recovering the complete lattice structure topology.

[0071] The restored lattice structure features are then input into a mechanical performance surrogate model that has undergone parameter freezing for prediction, yielding predicted performance results. In this embodiment, the predicted performance results include at least the predicted peak stress and the predicted buffer energy absorption.

[0072] Compared to traditional methods that re-execute finite element analysis for each structural update, this embodiment replaces time-consuming simulations with surrogate models, enabling reinforcement learning to achieve high-frequency, low-cost policy iteration during the training phase. For example, in traditional finite element analysis, a single structural performance evaluation may take several minutes to several hours, while a surrogate model can complete performance prediction in just milliseconds, allowing reinforcement learning to complete a large number of trial-and-error explorations within an acceptable timeframe.

[0073] The beneficial effect of the above process is that by establishing a closed-loop execution path between the potential action vector, the decoding structure, and the surrogate model prediction, real-time physical feedback in the reinforcement learning process is realized. This allows the reinforcement learning agent to obtain reliable performance evaluation results without frequently calling high-cost finite element analysis, thereby significantly improving the efficiency of structural inverse design and ensuring that the generated structure is always optimized around the target performance conditions.

[0074] S107: Based on the deviation between the predicted performance result and the target performance condition, a reward feedback is constructed to drive the reinforcement learning agent to update the policy until the preset convergence condition is met.

[0075] Specifically, the reward feedback is constructed based on the deviation between the predicted performance result and the target performance condition, including: decoding the latent action vector output by the reinforcement learning agent into a conditional variational autoencoder model, and obtaining the predicted performance result through a mechanical performance proxy model; calculating the performance deviation between the predicted performance result and the target performance condition, including: if the predicted performance result falls outside a preset interval, the performance deviation is calculated using the preset interval boundary closest to the predicted performance result as the target performance condition; if the predicted performance result falls within the preset interval, the constraint deviation is determined to be zero; and constructing a composite reward function based on the performance deviation, including a precision guidance term, a performance optimization term, and a constraint penalty term, to drive the reinforcement learning agent to update its policy.

[0076] Among them, "reward feedback" refers to the evaluation signal given by the reinforcement learning environment based on the current action result of the agent, which guides the agent to gradually learn a better structure generation strategy; "strategy update" refers to the agent continuously correcting the action output pattern based on the reward result, so that the generated structure gradually approaches the target performance condition.

[0077] In this embodiment, reference Figure 7 A composite reward function is constructed, combining a nested proxy model and a generative model, to achieve the closed-loop inverse design of the lattice structure. Specifically, the composite reward function is: ; in, This represents the normalized target performance vector; the first term is the accuracy guide term based on a Gaussian distribution. For the precision guide term based on Gaussian distribution, The first term is the sensitivity tolerance bandwidth; the second term is the performance-oriented optimization term. To add optimization weights, The first term is a specific normalized performance index for which additional optimization is desired. This term is mainly used to provide a slight tendency to explore the best possible outcome while meeting hard design constraints; the third term is... Constraints and penalties: To punish weights, utilize The norm is used to calculate the absolute deviation, providing the agent with a globally dense gradient and penalizing actions that deviate from the target.

[0078] The "Gaussian distribution precision guidance term" is used to establish a local high-reward region near the target performance to enhance the agent's search ability in the neighborhood of the target performance. The "performance-oriented optimization term" is used to guide the agent to further search for better structural solutions while meeting the main performance objectives. The "constraint penalty term" is used to penalize actions that deviate from the target performance conditions. It calculates the deviation norm between the predicted performance result and the target performance condition, and gives lower rewards to actions that are far from the target region. The target performance condition can be a preset interval. When the predicted performance result falls outside the preset interval, the boundary of the preset interval closest to the predicted performance result is taken as the target performance condition, and the difference between the two is taken as the performance deviation. If the predicted performance result falls within the preset interval, it is determined that the predicted performance result meets the interval constraint, that is, the constraint deviation is zero.

[0079] Furthermore, this embodiment can support a variety of different inverse design patterns.

[0080] In the single-objective precise reverse engineering generation mode, a specific engineering requirement is set as a task of precisely matching a single performance index. In the single-objective interval-constrained reverse engineering generation mode, a specific engineering requirement is set as a task of controlling a continuous performance interval. In the multi-objective collaborative precise reverse engineering generation mode, a specific engineering requirement is set as a task of simultaneously satisfying multiple performance indices.

[0081] The beneficial effect of the above process is that by constructing a composite reward function that integrates Gaussian precision guidance, performance optimization search, and bias penalty mechanism, the reinforcement learning agent can not only achieve accurate hits on target performance, but also take into account structural performance optimization and constraint satisfaction capabilities, thereby improving the adaptability and robustness of lattice structure inverse design.

[0082] S108 is a lattice structure scheme based on the output of the converged reinforcement learning agent satisfying the target performance conditions.

[0083] Among them, "policy update" refers to continuously adjusting the probability distribution of action output based on reward feedback, so that the agent gradually learns the optimal search policy; "convergence" means that the action results output by the reinforcement learning agent tend to be stable, and the prediction performance results can continuously meet the target performance requirements.

[0084] In this embodiment, the reinforcement learning agent adopts an Actor-Critic architecture consisting of an actor network and a critic network. The actor network generates an action distribution based on the state space input, while the critic network evaluates the value of the current action in the target task, thus providing optimization direction for the actor network.

[0085] Furthermore, this embodiment employs a proximal policy optimization algorithm to train the reinforcement learning agent. Here, "proximal policy optimization algorithm" refers to a reinforcement learning optimization method based on probability ratio constraints. Its core function is to limit the magnitude of a single policy update to avoid policy mutations that could lead to training instability.

[0086] Specifically, the proximal policy optimization algorithm constructs a policy optimization objective function with a truncation mechanism based on the probability ratio before and after the policy update. By limiting the magnitude of a single policy update, it ensures stable training of the reinforcement learning agent in a continuous latent feature manifold. For example, when an update causes the agent's output action to change too much, the truncation mechanism automatically limits the update magnitude to avoid drastic oscillations during the latent space search process, thereby enhancing training stability.

[0087] After the reinforcement learning agent completes training and reaches the preset convergence condition, the target performance condition is input into the converged reinforcement learning agent. To avoid pattern collapse in the generated results, this embodiment further performs random sampling based on the action distribution to obtain multiple sets of potential action vectors.

[0088] "Random sampling" refers to using the action probability distribution output by the actor network to generate multiple differentiated action solutions in the latent space.

[0089] Subsequently, multiple sets of potential action vectors are input into the conditional decoder for decoding, and the corresponding digital topology encoding and three-dimensional dot matrix structure model are recovered. Finally, multiple dot matrix structure schemes that meet the target performance conditions and have different topologies are output.

[0090] For example, under the same performance target, the system can simultaneously generate multiple heterogeneous lattice structures with different path connectivity, different node layouts, and different rod distribution forms, so that engineers can further select the optimal solution based on manufacturing constraints, cost conditions, or space size requirements.

[0091] The beneficial effects of the above process are that by introducing the Actor-Critic reinforcement learning architecture and the near-end policy optimization algorithm, stable training of reinforcement learning agents in a continuous latent space is achieved. Furthermore, by using a random sampling mechanism, the limitations of traditional one-to-one structure mapping in inverse design are overcome, enabling the system to stably output multiple lattice structure schemes with significant topological differences under the premise of satisfying the same performance constraints. This improves the design flexibility and scheme redundancy in complex engineering scenarios, and ultimately realizes closed-loop generative inverse design of lattice structures oriented towards target performance conditions.

[0092] In another embodiment, the effectiveness of the above method was verified. First, a single-target precise reverse generation mode was executed. A specific engineering requirement was set as a single-target precise performance hit task, and the target peak stress was set. The target pressure is 4750 MPa. The state encoding logic is invoked to transform this target value into a normalized state vector. Input is given to the agent. The Gaussian precise guide term in the composite reward function is activated by adjusting the weighting coefficients. Increasing the reward gain in the neighborhood of the target value drives the agent to perform target optimization sampling in a 16-dimensional latent manifold to obtain the optimal action vector. . The action vector The data is restored to digital topology encoding via a decoding module, and a parametric modeling script is automatically triggered to generate the corresponding 3D lattice geometric model. The performance of the generated geometric model is verified using a finite element simulation process. The results of a generative inverse design case for achieving precise single-objective optimization with a peak stress of 4750 MPa are as follows: Figure 8 As shown.

[0093] In this embodiment, the actual finite element verification stress values ​​of the generated lattice structure under the target of 4750 MPa are 4884.46 MPa and 4861.96 MPa, respectively, with the relative error controlled within 5%, achieving high-precision closed-loop matching for specific mechanical indicators.

[0094] Execute the single-objective interval constraint inverse generation mode. Set specific engineering requirements as continuous performance interval control tasks, and set target buffer energy absorption. The allowable range is 4500 mJ to 5000 mJ. The state encoding logic is invoked to transform the center value or boundary feature corresponding to this performance range into a state vector. Activating the compound reward function Constraint penalty terms, by adjusting the penalty weight coefficient For predicted performance values Action vectors falling outside the defined interval are penalized with high gain, forcing the agent's policy distribution to concentrate towards the target performance band region. During convergence, the agent outputs multiple sets of heterogeneous action vectors that satisfy the interval constraints through random exploration within the 16-dimensional latent manifold. Multiple sets of action vectors are reconstructed into a lattice geometric model via a decoding module and a modeling script, and then finite element verification is performed. The results of a generative inverse design case for achieving single-objective continuous-range control of energy absorption between 4500-5000 mJ are as follows: Figure 9 As shown.

[0095] In this embodiment, the prediction errors of the generated typical structures are 1.28%, 0.39% and 0.20%, respectively, and the measured buffer energy absorption of all structures strictly falls within the range of 4500-5000 mJ, realizing the automatic synthesis of diverse topologies under the constraints of nonlinear mechanical range; Execute a multi-objective collaborative precise inverse generation mode. Define specific engineering requirements as tasks where multiple performance indicators are simultaneously satisfied, and establish a dual-objective performance vector. The target peak stress is 4500 MPa, and the target buffer energy absorption is 7000 mJ. State encoding logic is invoked to transform the dual-objective performance requirements into a multi-dimensional state vector. Input is given to the agent. The multi-objective trade-off mechanism in the composite reward function is activated by collaboratively adjusting the weights of the accuracy-guided term. Weight of penalty items Calculate the predicted performance vector With dual target vectors The L2 distance between the two, i.e., the Euclidean distance error, serves as a comprehensive feedback signal to drive policy updates. The agent identifies Pareto-consistent regions satisfying dual physical constraints within a 16-dimensional continuous latent manifold and outputs the corresponding action vectors. The generated motion vectors are reconstructed via a decoding module and verified using a finite element simulation process. The results of a generative inverse design case for achieving precise multi-objective control with a peak stress of 4500 MPa and energy absorption of 7000 mJ are as follows: Figure 10 As shown.

[0096] In this embodiment, the peak stress error of the generated typical structure under dual-objective constraints is as low as 1.92% and 1.48%, respectively, and the buffer energy absorption error is as low as 0.34% and 1.77%, respectively. Moreover, under the same performance constraints, it outputs heterogeneous lattice schemes with significant topological differences, realizing multi-objective high-precision collaborative design under strong coupling constraints.

[0097] Based on the same inventive concept, this application also provides a lattice structure closed-loop generative inverse design apparatus for implementing the aforementioned lattice structure closed-loop generative inverse design method. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more embodiments of the lattice structure closed-loop generative inverse design apparatus provided below can be found in the limitations of the lattice structure closed-loop generative inverse design method described above, and will not be repeated here.

[0098] In one embodiment, such as Figure 11 As shown, a closed-loop generative inverse design device for lattice structures is provided, the device comprising: The vector establishment module 30 is used to obtain the target performance conditions of the lattice structure and establish a digital topology encoding vector to describe the topology of the lattice structure. The mechanical simulation module 31 is used to construct a parameterized simulation process for lattice structures based on digital topology encoding vectors, perform mechanical simulations on multiple topology encoding samples, extract corresponding mechanical performance indicators, and construct a mapping dataset between the topological features of lattice structures and mechanical performance indicators. Model building module 32 is used to build a mechanical performance proxy model based on the mapping dataset to achieve rapid prediction of the mechanical performance of lattice structures; The feature mapping module 33 is used to construct a conditional variational autoencoder model constrained by the target performance conditions, map the lattice structure features to a continuous latent feature manifold, and reconstruct the latent vectors to lattice structure features through the decoder; The reinforcement learning module 34 is used to construct a reinforcement learning agent, taking the target performance conditions as the input to the state space and the continuous latent feature manifold as the action space, and using the reinforcement learning agent to output the latent action vector. The result prediction module 35 is used to decode the latent action vector input conditional variational autoencoder model to obtain the reconstructed lattice structure features, and to predict the mechanical properties of the reconstructed lattice structure features through a mechanical performance surrogate model to obtain the prediction performance results. The policy update module 36 is used to construct reward feedback based on the deviation between the predicted performance result and the target performance condition, so as to drive the reinforcement learning agent to update the policy until the preset convergence condition is met. The scheme output module 37 is used to output a lattice structure scheme that satisfies the target performance conditions based on the converged reinforcement learning agent.

[0099] This application also provides an electronic device, in some embodiments, referring to... Figure 12 As shown, the electronic device 700 includes an input unit 710, a memory 720, a processor 730, and an output unit 740. The memory 720 stores program instructions that can be executed on the processor 730. The processor 730 can execute the closed-loop generative inverse design method and / or technical solution based on the lattice structure in the foregoing embodiments by calling the program instructions. The electronic device 700 can be a mobile terminal device such as a mobile phone or a computer.

[0100] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0101] The technical features of the above embodiments can be arbitrarily integrated. For the sake of brevity, not all possible integrations of the technical features in the above embodiments are described. However, as long as the integration of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0102] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A closed-loop generative inverse design method for lattice structures, characterized in that, The method includes: Obtain the target performance conditions of the lattice structure and establish a digital topology encoding vector to describe the topology of the lattice structure; Based on the digital topology encoding vector, a parameterized simulation process for lattice structures is constructed. Mechanical simulation is performed on multiple topology encoding samples to extract the corresponding mechanical performance indicators and construct a mapping dataset between the topological features of the lattice structure and the mechanical performance indicators. A mechanical performance proxy model is constructed based on the mapping dataset to achieve rapid prediction of the mechanical performance of lattice structures; A conditional variational autoencoder model constrained by target performance conditions is constructed, which maps lattice structure features to continuous latent feature manifolds, and reconstructs latent vectors into lattice structure features through a decoder; Construct a reinforcement learning agent, take the target performance conditions as the state space input, take the continuous latent feature manifold as the action space, and use the reinforcement learning agent to output the latent action vector. The latent action vector is input into the conditional variational autoencoder model for decoding to obtain the reconstructed lattice structure features. The mechanical performance of the reconstructed lattice structure features is then predicted using the mechanical performance proxy model to obtain the predicted performance results. Reward feedback is constructed based on the deviation between the predicted performance results and the target performance conditions to drive the reinforcement learning agent to update the policy until the preset convergence condition is met. A lattice structure scheme based on the output of the converged reinforcement learning agent that satisfies the target performance conditions.

2. The closed-loop generative inverse design method for lattice structures according to claim 1, characterized in that, The establishment of a digital topology encoding vector for describing the topology of the dot matrix structure includes: For multiple independent topological paths in the dot matrix structure, separate path control subsets are set. Each path control subset includes connection mode variables and node selection variables. The connection mode variable is used to represent the connection form between adjacent nodes, and the node selection variable is used to represent the spatial traversal relationship of the topological path in the candidate nodes, so as to form the digital topology encoding vector. The dot matrix structure is parametrically reconstructed based on the digital topology encoding vector to obtain the corresponding three-dimensional dot matrix structure model.

3. The closed-loop generative inverse design method for lattice structures according to claim 1 or 2, characterized in that, The dataset used to construct the mapping between the topological features of the lattice structure and its mechanical performance indicators includes: The digitized topology-coded vector is subjected to hierarchical sampling to generate multiple topology-coded samples; The parametric simulation process is invoked to perform compression simulation on the topology-coded sample and extract the compression response curve. Peak stress and energy absorption indices are extracted based on the compression response curve, and the mechanical performance indices are paired with corresponding topologically encoded samples to form a mapping dataset.

4. The closed-loop generative inverse design method for lattice structures according to claim 1, characterized in that, The construction of the mechanical performance proxy model includes: Perform physical feature engineering transformation on the digitized topology encoding vector to obtain a continuous physical feature vector; The continuous physical feature vector is input into a deep residual neural network to extract the coupling features between the lattice structure and mechanical properties. The deep residual neural network is iteratively trained based on the mechanical performance indicators in the mapping dataset, and the network parameters are frozen after the preset performance evaluation indicators are met, so as to form the mechanical performance proxy model.

5. The closed-loop generative inverse design method for lattice structures according to claim 4, characterized in that, The construction of the conditional variational autoencoder model constrained by the target performance conditions includes: The continuous physical feature vector is concatenated with the target performance condition and input into the probabilistic encoder to output the mean vector and variance parameter in the latent space. Reparameterized sampling is performed based on the mean vector, variance parameter, and random noise to obtain the latent vector; The latent vector and target performance conditions are input into the decoder for feature reconstruction. A composite loss function is constructed, comprising a geometric reconstruction loss term, a distribution regularization divergence term, and a physical consistency constraint loss term; the composite loss function is as follows: ； in, For geometric reconstruction loss term, Here, β is the KL divergence term, and β is the balance weight coefficient. For the physical consistency constraint loss term, As a surrogate model for mechanical properties, The feature vector reconstructed by the decoder. Let be the given condition vector; λ be the penalty weight coefficient. In the physical consistency constraint loss term, the mechanical performance proxy model is invoked to perform performance consistency verification on the reconstruction results, so as to form a continuous latent feature manifold that satisfies the physical constraints.

6. The closed-loop generative inverse design method for lattice structures according to claim 1 or 5, characterized in that, The method of constructing reward feedback based on the deviation between the predicted performance results and the target performance conditions includes: The latent action vectors output by the reinforcement learning agent are input into the conditional variational autoencoder model for decoding, and the prediction performance results are obtained through the mechanical performance proxy model. Calculate the performance deviation between the predicted performance result and the target performance condition; including: if the predicted performance result falls outside a preset interval, then the performance deviation is calculated using the preset interval boundary closest to the predicted performance result as the target performance condition; if the predicted performance result falls within the preset interval, then the constraint deviation is determined to be zero. Based on the performance deviation, a composite reward function is constructed, comprising a precision guidance term, a performance optimization term, and a constraint penalty term, to drive the reinforcement learning agent to update its policy; the composite reward function is as follows: ； in, This represents the normalized target performance vector; the first term is the accuracy guide term based on a Gaussian distribution. For the precision guide term based on Gaussian distribution, The first term is the sensitivity tolerance bandwidth; the second term is the performance-oriented optimization term. To add optimization weights, The first term is a specific normalized performance index for which additional optimization is desired. This term is mainly used to provide a slight tendency to explore the best possible outcome while meeting hard design constraints; the third term is... Constraints and penalties: To punish weights, utilize The norm is used to calculate the absolute deviation, providing the agent with a globally dense gradient and penalizing actions that deviate from the target.

7. The closed-loop generative inverse design method for lattice structures according to claim 6, characterized in that, The reinforcement learning agent adopts an Actor-Critic architecture consisting of an actor network and a critic network, and uses a proximal policy optimization algorithm to update the policy. By limiting the probability ratio before and after the policy update, the reinforcement learning agent can be stably trained in a continuous latent feature manifold.

8. The closed-loop generative inverse design method for lattice structures according to claim 1, characterized in that, The lattice structure scheme based on the converged reinforcement learning agent output satisfying the target performance condition includes: Input the target performance conditions into the converged reinforcement learning agent; Random sampling is performed based on the action distribution output by the reinforcement learning agent to obtain multiple sets of potential action vectors; The multiple sets of potential action vectors are decoded to output multiple lattice structure schemes that meet the target performance conditions and have different topologies.

9. A lattice structure closed-loop generative inverse design device, characterized in that, The device includes: The vector creation module is used to obtain the target performance conditions of the lattice structure and to create a digital topology encoding vector to describe the topology of the lattice structure. The mechanical simulation module is used to construct a parameterized simulation process for lattice structures based on the digital topology encoding vector, perform mechanical simulations on multiple topology encoding samples, extract the corresponding mechanical performance indicators, and construct a mapping dataset between the topological features of the lattice structure and the mechanical performance indicators. The model building module is used to build a mechanical performance proxy model based on the mapping dataset to achieve rapid prediction of the mechanical performance of the lattice structure. The feature mapping module is used to construct a conditional variational autoencoder model constrained by the target performance conditions, map the lattice structure features to a continuous latent feature manifold, and reconstruct the latent vectors to lattice structure features through the decoder; The reinforcement learning module is used to construct a reinforcement learning agent, taking the target performance conditions as the state space input and the continuous latent feature manifold as the action space, and using the reinforcement learning agent to output the latent action vector. The result prediction module is used to input the potential action vector into the conditional variational autoencoder model for decoding to obtain the reconstructed lattice structure features, and to predict the mechanical properties of the reconstructed lattice structure features through the mechanical performance proxy model to obtain the predicted performance results. The policy update module is used to construct reward feedback based on the deviation between the predicted performance result and the target performance condition, so as to drive the reinforcement learning agent to update the policy until the preset convergence condition is met. The scheme output module is used to output a lattice structure scheme that meets the target performance conditions based on the converged reinforcement learning agent.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the closed-loop generative inverse design method for lattice structures as described in any one of claims 1 to 8.

Images