A heterogeneous lattice structure generative inverse design method, device and equipment

Abstract

The application discloses a heterogeneous lattice structure generation type reverse design method, device and equipment, and relates to the technical field of intelligent structure design. The method comprises the following steps: establishing a digital design space of a heterogeneous lattice structure, and constructing a topological coding vector; obtaining mechanical property data of the heterogeneous lattice structure under multiple working conditions and multiple directions, and establishing training samples; training multiple mechanical property proxy models corresponding to different working conditions and loading directions; constructing a generation process of the heterogeneous lattice structure as a sequential decision process, and taking the proxy model as a performance evaluator in a reinforcement learning environment to perform strategy training; generating multiple candidate topological coding vectors by using the generated strategy after the training is completed, and screening to obtain heterogeneous lattice structure design schemes that meet target mechanical property requirements and are different in topological structure. The application can realize automatic reverse design of the heterogeneous lattice structure, improve design efficiency, and simultaneously generate multiple optional design schemes that meet target performance requirements while taking into account multiple performance constraints.

Description

Technical Field

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

[0002] As a typical artificially designed mechanical metamaterial, lattice structures have been widely used in aerospace, transportation, impact protection, biomedicine, 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 fabrication capability of complex topological lattice structures is continuously improving, providing greater freedom for the design of mechanical metamaterials.

[0003] Research on lattice structures can be broadly divided into experience-based forward design and performance-driven reverse design. Performance-driven reverse design, which involves directly solving for structural configurations that meet pre-defined mechanical performance targets, has become a research hotspot in this field. Reverse design, as a more advanced method, can design structures that satisfy desired performance targets. However, due to its extremely rich possibilities, reverse design of lattice structures remains challenging.

[0004] In recent years, data-driven methods have significantly improved the efficiency and accuracy of reverse engineering, such as the use of dual-model neural networks, convolutional neural networks combined with genetic algorithms, conditional generative adversarial networks, and physics-guided dual-network models. However, existing research largely focuses on isomorphic lattice structures, where the entire structure is designed by periodically arranging single-type unit cells, and performance targets are mostly limited to mechanical responses under a single direction or load condition. Regarding the customization of anisotropic performance in lattice structures, existing work is mostly still within the isomorphic framework, achieving limited anisotropic control by adjusting member diameter ratios, member hollowness and cross-sectional shape, shell thickness distribution, or introducing geometric twists. Heterogeneous lattices, by combining cells with different dominant mechanical behaviors, can achieve heterogeneous distributions of mechanical properties, thus allowing for more flexible control of anisotropic responses and possessing the potential to simultaneously satisfy constraints on multiple directions, multiple operating conditions, and anisotropic ratios.

[0005] Based on the above background, this invention proposes a reverse design method for heterogeneous lattice structures based on deep reinforcement learning. By integrating a surrogate model, it achieves high-precision, diversified, and automated reverse generation for multiple directions, multiple working conditions, and anisotropic constraints, and also achieves diversified topological structures in the design results. Summary of the Invention

[0006] The purpose of this invention is to provide a generative reverse design method, apparatus, and device for heterogeneous lattice structures. By establishing a digital design space for heterogeneous lattice structures, a unified encoding expression of the heterogeneous lattice structure topology is achieved. Furthermore, a surrogate model is established using multi-condition and multi-directional mechanical performance data. This surrogate model replaces finite element analysis as the performance evaluator in the reinforcement learning environment, significantly reducing the computational overhead during reinforcement learning training while ensuring the accuracy of mechanical performance prediction, thus improving the efficiency of heterogeneous lattice structure reverse design. Simultaneously, the heterogeneous lattice structure generation process is constructed as a sequential decision-making process, and a reward mechanism is built in conjunction with the target mechanical performance deviation. This allows the generation strategy to simultaneously consider multiple conditions such as compression, bending, and torsion, as well as anisotropic constraints, achieving automatic reverse design oriented towards target performance. Further, multiple heterogeneous lattice structure design schemes with different topologies but satisfying the same target mechanical performance requirements are generated through probabilistic sampling, improving the diversity of design results and the flexibility of engineering applications. This provides an efficient and accurate generative reverse design method for the intelligent design of complex heterogeneous lattice structures.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a generative reverse design method for heterogeneous lattice structures, the method comprising: A digital design space for heterogeneous lattice structures is established, each basic unit cell that makes up the heterogeneous lattice structure is discretely encoded, and a topological encoding vector is constructed based on the spatial position of each basic unit cell in the heterogeneous lattice structure, so that the topological encoding vector and the heterogeneous lattice structure are uniquely corresponded. Based on the topology encoding vector, heterogeneous lattice structure samples are constructed, and the mechanical performance data of each heterogeneous lattice structure sample under multiple loading conditions and multiple loading directions are obtained to construct training samples between the topology encoding vector and the mechanical performance data. Based on training samples, multiple mechanical performance proxy models are established, each corresponding to different loading conditions and loading directions. These mechanical performance proxy models can predict the target mechanical performance of the corresponding heterogeneous lattice structure based on the input topological encoding vector. The generation process of the heterogeneous lattice structure is constructed as a sequential decision-making process that determines the type of each spatial cell according to a preset cell filling order. The topological encoding vector of the partially filled cell and the current position to be filled are used as the state of the reinforcement learning agent, the cell type corresponding to the current position to be filled is used as the action, and the mechanical performance proxy model is used as the performance evaluator in the reinforcement learning environment. The preset cell filling order is one of a variety of candidate filling orders, and different filling orders correspond to different spatial position traversal paths. Reward information is constructed based on the deviation between the target mechanical performance predicted by the mechanical performance proxy model and the preset target mechanical performance. The reward information is used to train the reinforcement learning agent to obtain the heterogeneous lattice structure generation strategy. Multiple candidate topology encoding vectors are generated using the trained heterogeneous lattice structure generation strategy. The target mechanical properties of the heterogeneous lattice structure corresponding to each candidate topology encoding vector are predicted by a mechanical performance proxy model. Multiple heterogeneous lattice structure design schemes that meet the target mechanical performance requirements and have different topologies are selected.

[0008] Secondly, the present invention also provides a heterogeneous lattice structure generative reverse design apparatus, the apparatus comprising: The vector establishment module is used to establish a digital design space for heterogeneous lattice structures. It performs discrete encoding on each basic unit cell that makes up the heterogeneous lattice structure and constructs a topological encoding vector based on the spatial position of each basic unit cell in the heterogeneous lattice structure, so that the topological encoding vector and the heterogeneous lattice structure are uniquely corresponded. The sample construction module is used to construct heterogeneous lattice structure samples based on topology encoding vectors, and to obtain the mechanical performance data of each heterogeneous lattice structure sample under multiple loading conditions and multiple loading directions, so as to construct training samples between topology encoding vectors and mechanical performance data. The model building module is used to build multiple mechanical performance proxy models corresponding to different loading conditions and different loading directions based on training samples, so that the mechanical performance proxy models can predict the target mechanical performance of the corresponding heterogeneous lattice structure according to the input topological encoding vector. The process construction module is used to construct the generation process of the heterogeneous lattice structure as a sequential decision-making process that determines the type of each spatial cell according to a preset cell filling order. The topological encoding vector of the partially filled cell and the current position to be filled are used as the state of the reinforcement learning agent, the cell type corresponding to the current position to be filled is used as the action, and the mechanical performance proxy model is used as the performance evaluator in the reinforcement learning environment. The preset cell filling order is one of a variety of candidate filling orders, and different filling orders correspond to different spatial position traversal paths. The strategy acquisition module is used to construct reward information based on the deviation between the target mechanical performance predicted by the mechanical performance proxy model and the preset target mechanical performance. The reward information is used to train the reinforcement learning agent to obtain the heterogeneous lattice structure generation strategy. The performance prediction module is used to generate multiple candidate topology encoding vectors using the trained heterogeneous lattice structure generation strategy, predict the target mechanical performance of the heterogeneous lattice structure corresponding to each candidate topology encoding vector through a mechanical performance proxy model, and select multiple heterogeneous lattice structure design schemes that meet the target mechanical performance requirements and have different topologies.

[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 heterogeneous lattice structure generative reverse design method provided in the first aspect.

[0010] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the heterogeneous lattice structure generative reverse design method provided in the first aspect.

[0011] The beneficial effects of this invention are as follows: By establishing a digital design space for heterogeneous lattice structures, this invention achieves a unified encoding expression of the topology of heterogeneous lattice structures. Furthermore, by combining multi-condition and multi-directional mechanical performance data to establish a surrogate model, this surrogate model replaces finite element analysis as the performance evaluator in the reinforcement learning environment. While ensuring the accuracy of mechanical performance prediction, this significantly reduces the computational overhead during reinforcement learning training, thereby improving the efficiency of reverse design of heterogeneous lattice structures. Simultaneously, the heterogeneous lattice structure generation process is constructed as a sequential decision-making process, and a reward mechanism is built in conjunction with the target mechanical performance deviation. This enables the generation strategy to simultaneously consider multiple conditions such as compression, bending, and torsion, as well as anisotropic constraints, achieving automatic reverse design oriented towards target performance. Furthermore, by generating multiple heterogeneous lattice structure design schemes that meet the same target mechanical performance requirements but have different topologies through probabilistic sampling, this improves the diversity of design results and the flexibility of engineering applications, providing an efficient and accurate generative reverse design method for the intelligent design of complex heterogeneous lattice structures.

[0012] 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

[0013] Figure 1 This is a flowchart illustrating a heterogeneous lattice structure generative reverse design method according to an embodiment of the present invention. Figure 2 This is a diagram illustrating the encoding sequence of a heterogeneous lattice structure according to an embodiment of the present invention; Figure 3 This is a compressed model diagram of a heterogeneous lattice structure according to an embodiment of the present invention; Figure 4 This is a torsion model diagram of a heterogeneous lattice structure according to an embodiment of the present invention; Figure 5 This is a bending model diagram of a heterogeneous lattice structure according to an embodiment of the present invention; Figure 6This is a diagram of a deep feedforward neural network architecture for a proxy model according to an embodiment of the present invention; Figure 7 This is a schematic diagram of a Dueling DQN network architecture with unit encoding embedding and position index embedding according to an embodiment of the present invention; Figure 8 This is a diagram illustrating the filling order of three preset sequences according to an embodiment of the present invention; Figure 9 This is a reverse generation result diagram of a lattice structure guided by a single-condition, single-direction performance target, as shown in an embodiment of the present invention. Figure 10 This is a reverse generation result diagram of a lattice structure guided by multi-directional performance targets under a single working condition, as shown in an embodiment of the present invention. Figure 11 This is a reverse generation result diagram of a multi-condition performance target-oriented lattice structure according to an embodiment of the present invention; Figure 12 This is a diagram showing the reverse generation result of fusing anisotropic constraints according to an embodiment of the present invention. Detailed Implementation

[0014] 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.

[0015] 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.

[0016] 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.

[0017] Example 1 like Figure 1As shown, this embodiment provides a generative reverse design method for heterogeneous lattice structures. The reverse design of a heterogeneous lattice structure is completed through a process of constructing a digital design space, building training samples for a surrogate model, establishing a mechanical performance surrogate model, training a reinforcement learning strategy, and generating diverse design schemes. A heterogeneous lattice structure refers to a lattice structure formed by combining two or more basic cells with different topologies according to a preset spatial layout. Unlike traditional isomorphic lattice structures formed by periodically repeating the same type of cell, heterogeneous lattice structures can fully utilize the mechanical properties of different cells, thereby achieving synergistic optimization of comprehensive mechanical performance under multiple directions and working conditions. Generative reverse design refers to using the target mechanical performance as input and the structural topology as output, automatically searching for structural schemes that meet the target performance requirements through intelligent algorithms, without the need for repeated manual modification of structural parameters, thus achieving automatic mapping from target performance to structural topology.

[0018] S101. Establish a digital design space for the heterogeneous lattice structure, discretely encode each basic unit cell that makes up the heterogeneous lattice structure, and construct a topological encoding vector based on the spatial position of each basic unit cell in the heterogeneous lattice structure, so that the topological encoding vector establishes a unique correspondence with the heterogeneous lattice structure.

[0019] Heterogeneous lattice structures refer to lattice structures formed by combining multiple basic unit cells of different types in a certain spatial arrangement. Among them, the basic unit cell is the smallest repeating structural unit that constitutes the lattice structure, and different types of basic unit cells have different geometric topological forms and mechanical response characteristics.

[0020] In order to achieve the combined design of unit cells with different mechanical behaviors in heterogeneous lattice structures, this embodiment selects basic unit cells with different dominant mechanical behaviors as structural components.

[0021] Specifically, the basic unit cells include bending-dominant unit cells and tension-dominant unit cells. Among them, bending-dominant unit cells mainly bear external loads through the bending deformation of the members and have good deformation capacity; tension-dominant unit cells mainly bear external loads through the axial tension or compression of the members and have high load-bearing capacity.

[0022] This embodiment classifies lattice structures in three-dimensional space according to Maxwell's criterion. When the structural degrees of freedom meet the corresponding conditions, the structure is identified as a bending-dominant structure; when the structural degrees of freedom meet another condition, the structure is identified as a tension-dominant structure.

[0023] Maxwell's criterion classifies lattice structures in three-dimensional space as follows: when the degrees of freedom When it is a bending-dominant structure, It is a tension-dominated structure.

[0024] The above classification method can be used to determine the main mechanical behavior of the basic unit cell based on its own topological characteristics, providing a basis for the subsequent construction of heterogeneous lattice structures with different mechanical response characteristics.

[0025] Furthermore, this embodiment selects three basic unit cells that have a complete cubic shape, meet the requirements of additive manufacturing without significant suspended features, and possess inherent stability, namely BCCz unit cell, FCCz unit cell, and SC unit cell.

[0026] Among them, BCCz cells and SC cells belong to bending-dominated structures, while FCCz cells belong to stretching-dominated structures.

[0027] To facilitate computer description of heterogeneous lattice structures, this embodiment further performs discrete encoding on the aforementioned basic unit cell.

[0028] Specifically, each basic unit cell is assigned a unique discrete code value: BCCz unit cell corresponds to code 1; FCCz unit cell corresponds to code 2; and SC unit cell corresponds to code 3. This encoding method allows basic unit cells with different geometric structures to be represented by discrete numbers. For example, when a BCCz unit cell is placed at a certain spatial location, the corresponding code is 1; when an FCCz unit cell is placed, the corresponding code is 2; and when an SC unit cell is placed, the corresponding code is 3. After encoding, the geometric unit cell type information in complex 3D lattice structures is converted into discrete data, providing a unified data representation for subsequent topological encoding vector construction and machine learning model processing.

[0029] After completing the basic unit cell encoding, we further established a spatial layout representation of the heterogeneous lattice structure.

[0030] Spatial layout refers to the arrangement and positional relationship of different basic unit cells in three-dimensional space, and is used to describe the positional relationship between different types of unit cells.

[0031] In this embodiment, the target heterogeneous lattice structure is defined as a combination formed by topological interconnection of the above-mentioned basic unit cells at predetermined positions in three-dimensional space.

[0032] Specifically, based on the preset structural dimensions and spatial arrangement, the target structure is divided into multiple unit cell positions, each of which is used to place a basic unit cell. Subsequently, according to the preset spatial filling order, the basic unit cell type corresponding to each spatial position is determined sequentially.

[0033] like Figure 2As shown, this embodiment uses a preset encoding order to encode the heterogeneous lattice structure. Specifically, during the encoding process, each cell position is traversed according to a preset space-filling order, and the basic cell type code corresponding to each cell position is read. Multiple discrete codes are then arranged sequentially to form a one-dimensional discrete encoding vector. For example, when a heterogeneous lattice structure contains multiple cell positions, the basic cell type corresponding to the first spatial position is first determined, and the corresponding code is recorded. Subsequently, the basic cell types corresponding to the second, third, and so on, up to the last spatial position, are determined sequentially, ultimately forming a complete one-dimensional discrete encoding vector.

[0034] This embodiment adopts a cubic layout, including 8 unit cell positions, and follows the... Figure 2 The preset space filling order shown records the basic unit cell type codes at each spatial location sequentially, forming a one-dimensional discrete coding vector of length 8. There is a unique correspondence between the encoding vector and the heterogeneous lattice structure.

[0035] Specifically, for any given encoding vector, a heterogeneous lattice structure can be uniquely determined based on the basic unit cell types and their arrangement order corresponding to each element in the encoding vector. Conversely, for any heterogeneous lattice structure, a unique corresponding encoding vector can also be generated based on the basic unit cell type corresponding to its spatial location. Therefore, all possible encoding vectors together constitute the digital design space of heterogeneous lattice structures. In this digital design space, each encoding vector represents a candidate heterogeneous lattice structure design scheme, and different encoding vectors correspond to different unit cell combinations and spatial topological relationships.

[0036] By establishing the aforementioned digital design space, the traditional method of describing heterogeneous lattice structures based on geometric models is transformed into a data description method based on discrete encoded vectors, enabling complex heterogeneous lattice topologies to be directly stored, searched, and optimized by computers.

[0037] This embodiment encodes different types of basic cells and constructs topological encoding vectors by combining spatial layout relationships, realizing the digital expression of topological information of heterogeneous lattice structures. This provides a unified data foundation for subsequent performance data generation based on finite element simulation, surrogate model training, and reinforcement learning reverse generation.

[0038] S102, construct heterogeneous lattice structure samples based on topology encoding vectors, and obtain the mechanical performance data of each heterogeneous lattice structure sample under multiple loading conditions and multiple loading directions, so as to construct training samples between topology encoding vectors and mechanical performance data.

[0039] Training samples refer to data samples composed of topological information of heterogeneous lattice structures and corresponding mechanical performance information. Specifically, training samples include topological encoding vectors as input and mechanical performance data as output labels. By establishing the correspondence between the two, the computational model can learn the mapping relationship between the topology of heterogeneous lattice structures and their mechanical responses.

[0040] In the specific implementation process, the corresponding heterogeneous lattice structure model is first generated based on the topological encoding vector in the digital design space. Since the topological encoding vector only contains information about the basic unit cell type and its arrangement order, the discrete encoding needs to be further converted into a specific geometric structure.

[0041] This embodiment employs parametric modeling. Based on the basic unit cell type corresponding to each element in the topological encoding vector, it automatically calls the corresponding basic unit cell geometric model and completes the unit cell assembly according to a preset spatial layout to form a complete heterogeneous lattice structure model. Parametric modeling refers to setting design factors such as structural geometric dimensions, unit cell type, and arrangement as adjustable parameters, allowing for the automatic generation of different structural models by modifying these parameters. For example, when reading a topological encoding vector, if the first position of the vector corresponds to code 1, then the BCCz unit cell model is called to fill the first spatial position; if the second position corresponds to code 2, then the FCCz unit cell model is called to fill the second spatial position, and so on, ultimately forming a heterogeneous lattice structure uniquely corresponding to the topological encoding vector.

[0042] Furthermore, in order to achieve the automatic generation of a large number of samples with different topological structures, this embodiment adopts an automated finite element modeling method.

[0043] Specifically, a Python script is used to call finite element analysis software to batch model heterogeneous lattice structures corresponding to different topological encoding vectors. During the modeling process, the geometric model is determined based on the topological encoding vectors, material properties are set according to preset material parameters, a finite element mesh is generated according to the analysis requirements, and corresponding boundary conditions and loading methods are set, ultimately forming a parametric finite element model for mechanical property analysis. This method avoids manually building finite element models one by one, improving the efficiency of generating samples with different topological structures.

[0044] In order to enable the training samples to cover different structural forms in the digital design space, this embodiment further adopts the Latin hypercube sampling method to select structural samples from the digital design space of heterogeneous lattice structures.

[0045] Latin hypercube sampling is a spatially uniform sampling method that divides the range of design variables into multiple intervals and samples within each interval, so that a limited number of samples can cover the entire design space as much as possible.

[0046] Specifically, the position of each unit cell in the heterogeneous lattice structure is used as a design variable, and the basic unit cell type corresponding to each position is used as a discrete design parameter to sample the entire digital design space. Latin hypercube sampling is used to generate structural samples, and the formula corresponding to the continuous sampling process is as follows: ;in It is the final calculated sample value. The value range is 1 to , The total number of samples, Representing the There are design variables, assuming the structure contains... Each unit cell position, then The value range is 1 to The first term, 1, is the lower limit of the interval; the second term... Determine the main layer where the sample is located; Represents a random permutation; the third item It is a small random perturbation within the layer. It is an independent, uniformly distributed random number, ensuring that the sample points do not rigidly fall in the exact center of each layer. After obtaining the continuous sampling results, the continuous sampling results are discretized, and the continuous sampled values ​​are converted into the corresponding basic unit cell type codes, thereby obtaining a discrete topological encoding vector that can be used for finite element modeling.

[0047] Specifically, continuous sampling results are mapped to the basic unit cell coding range, and the corresponding discrete codes are obtained by rounding. For example, if a continuous sampling value corresponds to a certain basic unit cell coding interval after mapping, the value is converted into the corresponding integer code to form the final topological coding vector.

[0048] By using the above sampling method, representative heterogeneous lattice structure samples can be obtained under limited computing resources, thereby improving the coverage of training samples for different topological combinations.

[0049] After obtaining the finite element model of the heterogeneous lattice structure, this embodiment further establishes corresponding finite element analysis models for different loading conditions to obtain the corresponding mechanical response of the structure.

[0050] Specifically, finite element analysis models were established for compression, torsion, and bending conditions, respectively. For example... Figure 3 As shown, this embodiment establishes a finite element model of a heterogeneous lattice structure under compression loading. By applying a compressive load to the structure, the mechanical response of the structure under compression is obtained. In the compression condition, by setting a fixed end and a loading end, the heterogeneous lattice structure is subjected to compression along a preset direction, and the load changes and structural deformation generated during the loading process are recorded.

[0051] Furthermore, such as Figure 4 As shown, this embodiment establishes a finite element model of a heterogeneous lattice structure under torsional loading. By applying a torsional load, the mechanical response of the structure in resisting torsional deformation is obtained. Under torsional conditions, torsional constraints and rotational loads are set to induce rotational deformation of the structure around a preset axis, and corresponding torsional performance data are obtained based on the load changes.

[0052] Furthermore, such as Figure 5 As shown, this embodiment establishes a finite element model of a heterogeneous lattice structure under bending loading. By applying bending load, the mechanical response of the structure under bending is obtained. In the bending condition, by setting the loading position and constraint conditions, the structure is made to undergo bending deformation, and the corresponding mechanical response data is recorded.

[0053] To analyze the performance differences of the heterogeneous lattice structure in different directions, this embodiment simulates the structure along multiple loading directions or loading planes.

[0054] Specifically, under the same loading condition, the mechanical responses in different directions are obtained by changing the load direction or structural orientation. For example, under compression conditions, compressive loads can be applied along different spatial directions to analyze the differences in load-bearing capacity in different directions. After completing the finite element simulation, the output results of each finite element model are processed to obtain the corresponding mechanical performance data. This mechanical performance data describes the performance of the heterogeneous lattice structure under different loading conditions, including compressive bearing capacity, torsional torque, flexural moment, and anisotropy ratio.

[0055] Specifically, after the calculations are completed, an automated Python script reads the .odb files from the results database. For compression conditions, the support reactions and displacements at the reference points of the loading plate are extracted to generate force-displacement curves. The load at 10% of the engineering strain is taken as the compressive bearing capacity. And calculate the integral area under the curve before that point as the energy absorbed during compression. For torsional conditions, the reaction torque and rotation angle at the top reference point are extracted, and the torque at the specified rotation angle is taken as the anti-torsional torque. And integrate to obtain the torsional adsorption energy. For bending conditions, the bending moment is similarly obtained. and bending energy absorption Calculate the performance indices in the X, Y, and Z directions, or in the XZ, XY, and YZ planes, respectively: , , .

[0056] in, , , These represent the anisotropy ratios of the compression, torsion, and bending properties of the lattice structure, respectively. , , These represent the compressive bearing capacity of the lattice structure in the X, Y, and Z directions, respectively. , , These represent the torsional torque in the X, Y, and Z directions of the lattice structure, respectively. , , These represent the bending moments of the three planes XZ, XY, and YZ of the lattice structure, respectively.

[0057] Through the above calculations, we can obtain the multi-condition and multi-directional mechanical performance data corresponding to each heterogeneous lattice structure sample.

[0058] Subsequently, the topological encoding vectors are associated with the corresponding mechanical performance data to form training samples.

[0059] Specifically, for any training sample: the input part is the topological encoding vector corresponding to the heterogeneous lattice structure; the output part is the mechanical performance data of the structure under different loading conditions and different loading directions.

[0060] For example, a topology encoding vector corresponds to a heterogeneous lattice structure. The performance data of this structure in each direction under compression, torsion and bending conditions are obtained through finite element analysis. Then, the topology encoding vector is combined with the above performance data to form a complete training sample.

[0061] By constructing a large number of training samples, a data correlation can be established between the topological composition of heterogeneous lattice structures and their mechanical properties, providing a training data foundation for the subsequent establishment of mechanical performance proxy models.

[0062] This embodiment automatically generates heterogeneous lattice structure samples based on topological encoding vectors and obtains corresponding mechanical performance data through multi-condition and multi-directional finite element simulations. This achieves a unified expression of the topological information and mechanical response data of heterogeneous lattice structures, improves the efficiency of training data acquisition, and provides a reliable data foundation for subsequent data-driven structural performance prediction and reverse design.

[0063] S103. Based on the training samples, multiple mechanical performance proxy models are established, each corresponding to different loading conditions and different loading directions, so that the mechanical performance proxy models can predict the target mechanical performance of the corresponding heterogeneous lattice structure according to the input topological encoding vector.

[0064] Among them, the mechanical performance proxy model refers to a predictive model that learns the mapping relationship between the topological features of a heterogeneous lattice structure and its mechanical response through training data. It can replace part of the finite element simulation process and directly output the corresponding mechanical performance prediction results after inputting a new topological encoding vector.

[0065] Since the aforementioned topological encoding vector is composed of discrete unit cell type encoding, it is difficult to fully reflect the internal spatial correlation of the heterogeneous lattice structure when directly input into the model. Therefore, this embodiment first performs feature engineering processing on the topological encoding vector to convert the discrete encoding into a continuous feature vector.

[0066] Specifically, feature extraction is performed on the topological encoding vector corresponding to each heterogeneous lattice structure to obtain continuous features used to describe the topological information of the structure.

[0067] Continuous features include at least: basic normalized features, used to describe the cell encoding information after normalization; statistical features, used to describe the quantity and proportion of different types of basic cells in the overall structure; spatial correlation features, used to describe the arrangement relationship of basic cell types between different spatial locations; symmetry features, used to describe the spatial distribution pattern of the heterogeneous lattice structure along different directions; cell type distribution features, used to describe the distribution of different types of basic cells in the structure; and global aggregation features, used to characterize the topological composition characteristics of the heterogeneous lattice structure from an overall perspective.

[0068] The feature transformation process corresponds to the following feature engineering calculation method: (1) Basic normalization characteristics, i.e. ;in Indicates the first Unit cell type encoding for each spatial location, This indicates the position number of the unit cell in the filling order. The total number of basic unit cells in the heterogeneous lattice structure, i.e. the total number of spatial positions, is equal to the length of the encoding vector; (2) Statistical characteristics, including the mean ,variance Maximum and minimum values; (3) Spatial correlation characteristics, namely the product and absolute difference of adjacent position pairs, are defined as follows: , , ,in For the set of adjacent position encoding pairs, For two spatial location indices of adjacent location pairs, For adjacent position pairs Normalized encoded product, For adjacent position pairs The absolute difference in normalized encoding; (4) Symmetry measure, defined as ,in The set representing the centrally symmetric positional encoding pairs. These are two positional indices that belong to a centrally symmetric pairing; (5) Distribution frequency of unit cell types, i.e. Encode for all cell types, This indicates the frequency of occurrence of unit cell type k in the structure. For the indicator function when, The value is 1 if it is true, and 0 otherwise. (6) Global aggregation features, defined as The final D-dimensional feature vector is the concatenation of all the above components: ,in This is the final concatenated complete feature vector. Represents a D-dimensional real space. For all A normalized encoded value, It is the product of all adjacent pairs of positions. It is the absolute difference between all pairs of adjacent positions. The distribution frequencies of all unit cell types are then normalized using Z-score and used as input to the neural network.

[0069] After the above processing, the original discrete topological code is converted into a continuous feature vector, and the continuous feature vector is standardized to bring different types of features to a uniform data scale, thereby improving the stability of model training.

[0070] Furthermore, to address the issue of different mechanical responses in heterogeneous lattice structures under different loading conditions and directions, this embodiment establishes multiple mechanical performance proxy models.

[0071] Specifically, such as Figure 6 As shown, this embodiment uses a deep feedforward neural network to construct a mechanical performance proxy model. The deep feedforward neural network includes an input layer, a shared feature extraction layer, multiple task branch layers, and an output layer.

[0072] The input layer receives continuous feature vectors after feature engineering; the shared feature extraction layer extracts common topological features between different heterogeneous lattice structures, enabling the model to learn the correlation between basic unit cell combination, spatial layout relationship and mechanical performance; multiple task branch layers correspond to different prediction tasks, and are used to further predict the mechanical performance under different loading conditions and different loading directions based on shared features; the output layer outputs the final prediction result.

[0073] For example, a corresponding surrogate model is established for compression conditions to predict the load-bearing capacity of heterogeneous lattice structures under compression loading conditions; a corresponding surrogate model is established for torsion conditions to predict the torsional performance of the structure; and a corresponding surrogate model is established for bending conditions to predict the bending response of the structure.

[0074] During model training, the aforementioned training samples are input into a deep feedforward neural network. Continuous features corresponding to the topological encoding are used as input, and mechanical performance data obtained from finite element simulation are used as supervision labels to iteratively optimize the network parameters. During training, the network parameters are continuously adjusted by comparing the differences between the model's prediction results and the actual mechanical performance data. Once the model's prediction accuracy meets the preset requirements, training is stopped and the network parameters are fixed, forming the final mechanical performance proxy model. The trained mechanical performance proxy model can quickly predict the mechanical performance of the corresponding structure under the target loading condition and target loading direction based on the input heterogeneous lattice structure topological encoding vector. Compared to the need to rebuild and solve the finite element model for each candidate structure evaluation, this embodiment transforms the complex finite element calculation process into a fast prediction process by establishing a data-driven mechanical performance proxy model, thus improving the performance evaluation efficiency in the subsequent reinforcement learning structure search process.

[0075] This embodiment achieves rapid mapping between topological structure and mechanical performance by using multidimensional feature representation of heterogeneous lattice structure topology encoding and establishing mechanical performance proxy models for different loading conditions and loading directions. This provides a performance evaluation basis for subsequent automatic structure generation by reinforcement learning agents.

[0076] S104, the generation process of the heterogeneous lattice structure is constructed as a sequential decision-making process that determines the type of each spatial cell according to the preset cell filling order. The topological encoding vector of the partially filled cell and the current position to be filled are used as the state of the reinforcement learning agent, the cell type corresponding to the current position to be filled is used as the action, and the mechanical performance proxy model is used as the performance evaluator in the reinforcement learning environment.

[0077] The sequential decision-making process refers to breaking down the generation process of a complete heterogeneous lattice structure into multiple consecutive decision-making steps. In each decision-making step, a basic unit cell type corresponding to a spatial location is determined, and a complete topological structure is formed after multiple decisions.

[0078] Specifically, the unit cell filling order is pre-set based on the spatial layout relationship of the heterogeneous lattice structure.

[0079] like Figure 8As shown, this embodiment sets multiple cell filling orders, including different traversal methods and different spatial path filling methods. Since different filling orders can change the state transition trajectory, they may affect learning efficiency and policy convergence. To examine the robustness of the proposed framework to the decision order and verify the effectiveness of the position embedding mechanism, different filling orders are set. Specifically, three filling orders are preset: from bottom to top, filling the four positions at the bottom layer first, then filling the four positions at the top layer, i.e., following... Fill sequentially; from left to right, first fill the four positions on the left layer, then fill the four positions on the right layer, that is, according to... Fill sequentially; from front to back, first fill the four positions of the front layer, then fill the four positions of the back layer, that is, according to... Fill in the order.

[0080] During structure generation, each spatial location is visited sequentially according to a preset unit cell filling order, and the corresponding basic unit cell type is selected at each spatial location. To enable the reinforcement learning agent to perceive the current structure generation state, this embodiment defines state information during the reinforcement learning process. The state describes the structural information possessed by the reinforcement learning agent at the current decision-making moment, including partially determined topological information and the spatial location where a decision needs to be made.

[0081] Specifically, the partially filled topological encoding vector and the current position to be filled are used together as the state of the reinforcement learning agent. The partially filled topological encoding vector represents the encoding corresponding to the spatial position where the cell type selection has been completed; the current position to be filled represents the spatial position where the basic cell type needs to be determined in the next step. For example, in a heterogeneous lattice structure containing multiple cell positions, if the first few spatial positions have been filled, the current state includes the encoding information corresponding to the determined positions and the number of the currently unfilled position. The reinforcement learning agent uses this state to determine the current structure formation and selects the next filling method.

[0082] Furthermore, the basic cell type corresponding to the current position to be filled is used as the action of the reinforcement learning agent.

[0083] Here, "action" represents the structural adjustment behavior performed by the reinforcement learning agent in the current state. In this embodiment, each action corresponds to a basic unit cell type selection result.

[0084] For example, when the current position to be filled can be selected from three basic cells: BCCz, FCCz, and SC, the reinforcement learning agent has three candidate actions, representing the selection of a BCCz cell, an FCCz cell, or an SC cell to fill the current spatial position. After executing an action, the reinforcement learning agent writes the corresponding cell type encoding into the current topological encoding vector and proceeds to the decision-making process for the next spatial position until all spatial positions are filled.

[0085] In this way, the complex structural design problem that originally required determining all the combinations of individual cells at once is transformed into a sequential decision problem consisting of multiple consecutive steps, enabling the reinforcement learning agent to gradually optimize the structural topology.

[0086] Furthermore, the aforementioned mechanical performance proxy model is used as a performance evaluator in a reinforcement learning environment.

[0087] The reinforcement learning environment is used to provide feedback evaluation information based on the agent's current behavior, while the performance evaluator is used to quickly predict the mechanical properties of the generated structure.

[0088] Specifically, after the reinforcement learning agent completes a full topology generation, it obtains a complete topology encoding vector, which is then input into the mechanical performance proxy model. The mechanical performance proxy model predicts the mechanical performance of the corresponding heterogeneous lattice structure under the target loading condition and loading direction based on the input topology encoding vector, and uses the prediction results as environmental feedback information. Since the mechanical performance proxy model has already established a mapping relationship between the topology and mechanical performance, it is not necessary to re-execute a full finite element analysis to quickly obtain the performance evaluation results of the currently generated structure. During reinforcement learning training, the reinforcement learning agent continuously selects actions based on the current state and uses the mechanical performance proxy model to obtain environmental feedback, thereby gradually learning the influence of different unit cell combinations on structural performance.

[0089] This embodiment transforms the heterogeneous lattice structure generation process into a state- and action-based sequential decision-making process, and uses a mechanical performance proxy model to replace the traditional finite element repetitive calculation process, thereby achieving rapid performance evaluation during the structure generation process and providing a foundation for subsequent reinforcement learning strategy training.

[0090] S105, construct reward information based on the deviation between the target mechanical performance predicted by the mechanical performance proxy model and the preset target mechanical performance, and use the reward information to train the reinforcement learning agent to obtain the heterogeneous lattice structure generation strategy.

[0091] Among them, the generation strategy refers to the decision rule by which the reinforcement learning agent determines the next cell type selection method based on the current structural state, and it can output the corresponding structure generation action according to the target performance requirements.

[0092] Specifically, after the reinforcement learning agent completes a full structure generation, it obtains the corresponding complete topology encoding vector and inputs the topology encoding vector into the mechanical performance proxy model.

[0093] The mechanical performance proxy model outputs the predicted mechanical performance corresponding to the current heterogeneous lattice structure.

[0094] This embodiment constructs reward information based on the deviation between the predicted mechanical properties under the target working condition and the preset target mechanical properties.

[0095] The reward information is used to evaluate whether the current structure generation result meets the design goal, and serves as an important basis for the reinforcement learning agent to adjust the selection of subsequent actions.

[0096] Specifically, the predicted mechanical properties are compared with the target mechanical properties: When the predicted mechanical properties of the generated structure are close to the preset target value, it indicates that the current unit cell combination method can meet the design requirements, and the corresponding reward value is increased. When the predicted mechanical properties deviate significantly from the target value, it indicates that the current structural topology does not meet the design requirements, and the corresponding reward value is reduced.

[0097] Through the above methods, reinforcement learning agents can gradually tend to choose cell combination methods that can improve the degree of structural performance matching during the training process.

[0098] Furthermore, since heterogeneous lattice structures typically need to meet mechanical performance requirements in multiple directions simultaneously in practical applications, optimizing performance in only one direction may lead to a decrease in performance in other directions. Therefore, this embodiment further introduces anisotropic constraints.

[0099] Among them, anisotropic constraints are used to describe the degree of difference in mechanical properties of heterogeneous lattice structures under different loading directions.

[0100] Specifically, based on the performance data corresponding to multiple loading directions predicted by the mechanical performance proxy model, the differences between the performance in different directions are calculated, and the differences are used as constraints in the structure generation process.

[0101] For example, if a structure has a high load-bearing capacity in the X direction but its performance is significantly reduced in the Y and Z directions, it is considered to have a strong directional dependence and its reward rating should be reduced; if a structure maintains relatively balanced mechanical performance in multiple directions, its corresponding reward rating should be increased.

[0102] This embodiment uses the reward corresponding to the target performance deviation and the anisotropic constraint as feedback information of the reinforcement learning environment, so that the reinforcement learning agent not only focuses on the degree of target performance satisfaction, but also considers the overall performance balance of the structure.

[0103] The reward function can integrate performance indicators from multiple directions and operating conditions, as well as anisotropic constraints. ; in For the round reward function, , For the first One performance prediction value, For the target value, This is the weighting coefficient, usually set to 1. The total number of target performance metrics of interest; The anisotropy ratio is calculated based on multi-directional prediction values. The target is the anisotropy ratio; To prevent division by zero for small constants; The weights are for anisotropic constraints.

[0104] The trained proxy model is invoked to predict the load-bearing capacity or moment and energy absorption value of the structure under the target working condition and direction, and the terminal reward is calculated according to the following formula: .

[0105] After obtaining the reward information, this embodiment uses a reinforcement learning agent based on the Dueling DQN architecture for training.

[0106] like Figure 7 As shown, the Dueling DQN architecture includes a state input part, a feature extraction part, a state value branch, and an action advantage branch.

[0107] The state input part is used to receive the current heterogeneous lattice structure generation state; the feature extraction part is used to extract the structural features from the current part filling topology encoding vector and the information of the position to be filled; the state value branch is used to evaluate the overall value of the current state; and the action advantage branch is used to evaluate the degree of advantage of different basic units in choosing an action relative to other actions.

[0108] By combining state value and action advantage, the action value corresponding to different actions in the current state can be obtained.

[0109] Among them, action value is used to represent the probability of generating a heterogeneous lattice structure that meets the target mechanical performance requirements after selecting a certain basic unit cell type in the current state.

[0110] During the specific training process, the reinforcement learning agent outputs multiple candidate actions based on the current state and selects one action to execute. After the action is executed, the environment calls the mechanical performance proxy model to evaluate the performance based on the new structural state and returns the corresponding reward information. The reinforcement learning agent continuously updates the action value network parameters based on the current state, the executed action, the reward information, and the next state, enabling the network to gradually learn the correlation between actions generated by different structures and the final performance result.

[0111] To improve training stability, this embodiment further employs an experience playback mechanism.

[0112] Among them, the experience replay mechanism is used to store state transition data during the historical interaction process of the reinforcement learning agent.

[0113] Specifically, during training, data generated from each interaction is stored in an experience pool. This data includes: the current state; the current action; the reward received; and the next state after the action. When updating network parameters, multiple historical interaction samples are randomly selected from the experience pool for training. By randomly sampling historical experience, the correlation between consecutive training samples can be reduced, improving the stability of network parameter updates.

[0114] Furthermore, during training, the reinforcement learning strategy is updated based on the output of the action value network.

[0115] Specifically, the reinforcement learning agent continuously selects actions based on the current state and optimizes the action value estimation using reward information from environmental feedback, gradually increasing the selection probability of actions with higher performance gains. As the training process continues, the reinforcement learning agent can gradually learn the topological combination rules of heterogeneous lattice structures and form a structure generation strategy that satisfies the target mechanical performance constraints. When the training reaches the preset conditions, the reinforcement learning network parameters are fixed, and the trained heterogeneous lattice structure generation strategy is obtained. This generation strategy can automatically determine the basic unit cell type corresponding to each spatial location based on the input target performance requirements and the current structure generation state, providing a foundation for subsequently generating multiple different topological structure design schemes.

[0116] This embodiment constructs a reward mechanism that integrates target performance deviation and anisotropic constraints, and uses the Dueling DQN reinforcement learning framework to train the generation strategy, enabling the reinforcement learning agent to autonomously learn the relationship between the topology and mechanical properties of heterogeneous lattice structures, thereby improving the automation level of the reverse design process of complex heterogeneous lattice structures.

[0117] S106. Using the trained heterogeneous lattice structure generation strategy, multiple candidate topology encoding vectors are generated. The target mechanical performance of the heterogeneous lattice structure corresponding to each candidate topology encoding vector is predicted by the mechanical performance proxy model. Multiple heterogeneous lattice structure design schemes that meet the target mechanical performance requirements and have different topologies are selected.

[0118] Specifically, the heterogeneous lattice structure generation strategy after training is loaded into the structural design system, and the preset target mechanical performance requirements are input.

[0119] Among them, the preset target mechanical performance requirements are used to describe the expected performance indicators of the heterogeneous lattice structure, including target load-bearing capacity, target torsional resistance, target energy absorption capacity, and performance constraints under different loading directions.

[0120] During the generation process, the basic cell type corresponding to each spatial position is determined step by step according to the preset cell filling order.

[0121] Specifically, for the current location to be filled, the topological encoding vector of the current partial filling and the current location are input into the trained generation strategy. The generation strategy then outputs the action value corresponding to each candidate action in the current state. Here, the action value represents the probability of generating a structure that meets the target performance requirements after selecting a certain basic cell type in the current state. For example, for the current location to be filled, if the selectable basic cell types include BCCz, FCCz, and SC, the generation strategy calculates the action value corresponding to each of the three cell types and determines the contribution of different cell types to the current structure generation process based on the action value.

[0122] To improve the diversity of generated structures, this embodiment does not directly select the cell type with the highest action value, but determines the action probability based on the action value corresponding to different actions, and performs random sampling according to the action probability.

[0123] Specifically, the action value corresponding to each candidate action is converted into the action selection probability.

[0124] The probability of an action is calculated as follows: , in Let be the action value function, representing the state. Next action Then, the estimated value of the expected cumulative reward in the future. The denominator is the value of the candidate actions, and all candidate actions are iterated through. Calculate their respective Q values. To control the "smoothness" or "exploration intensity" coefficient of the probability distribution, typically... Take 1 to 1.2.

[0125] By adjusting the probability distribution, the balance between exploration capability and performance optimization capability during structure generation can be controlled. When more attention is paid to generating high-performance structures, the selection probability of high-value actions is increased, making the generation results more concentrated in the region with better performance. When more different topology design schemes are needed, the possibility of other candidate actions being selected is increased, enabling the generation process to explore more structural combinations.

[0126] Based on the aforementioned action probabilities, a basic unit cell type is randomly selected at the current spatial location, and its corresponding code is written into the topological coding vector. The current structural state is then updated, and the action selection process continues for the next spatial location until the unit cell type has been determined for all spatial locations. By repeating this process, multiple candidate topological coding vectors can be obtained.

[0127] Since the action sampling results differ in different generation processes, different candidate topological encoding vectors correspond to different basic unit cell space combinations, thus forming multiple heterogeneous lattice structure candidate schemes with different topological structures.

[0128] After obtaining multiple candidate topology encoding vectors, this embodiment further utilizes a mechanical performance surrogate model to perform rapid performance prediction on the candidate structures.

[0129] Specifically, each candidate topology encoding vector is input into the corresponding mechanical performance proxy model, and the proxy model predicts the mechanical performance of each candidate heterogeneous lattice structure under the target loading condition and the target loading direction.

[0130] The predicted performance data include: structural compressive strength, torsional strength, bending strength, structural energy absorption capacity, and calculated anisotropy ratio.

[0131] Since the mechanical performance proxy model has learned the mapping relationship between the topology and mechanical performance, it is possible to quickly evaluate multiple candidate structures without re-performing finite element modeling and simulation calculations.

[0132] Subsequently, candidate topology encoding vectors are filtered based on the prediction results.

[0133] Specifically, the predicted mechanical properties of each candidate structure are compared with the preset target mechanical property constraints. A candidate structure is retained if it meets all target performance requirements; otherwise, it is deleted. For example, if the target requirements are: the structure has a preset load-bearing capacity; the difference in load-bearing capacity in different directions meets the anisotropy constraint; and the energy absorption capacity reaches a preset range, then only heterogeneous lattice structures that simultaneously meet these conditions are retained. This screening process removes candidate structures with insufficient performance, improving the effectiveness of the final output design.

[0134] Furthermore, to avoid duplicate topological forms among multiple output structural schemes, this embodiment performs topological duplication detection on the candidate topology encoding vectors that meet performance constraints.

[0135] Among them, topological repetition detection is used to determine whether different candidate topological encoding vectors correspond to the same structural form.

[0136] Specifically, the basic cell codes corresponding to each spatial position in different candidate topological coding vectors are compared one by one.

[0137] If all spatial locations in two candidate topological encoding vectors have the same encoding, they are considered to correspond to the same heterogeneous lattice structure, and only one is retained. If at least one spatial location in two candidate topological encoding vectors has a different encoding, they are considered to have different topological structures, and both are retained. This method ensures that the final output of multiple heterogeneous lattice structure designs, while satisfying the same objective mechanical performance constraints, will have different topological compositions.

[0138] Finally, multiple heterogeneous lattice structure design schemes after performance screening and topology deduplication are output. The output results include: (1) the topology encoding vector corresponding to the heterogeneous lattice structure; (2) the structural model generated based on the topology encoding vector; and (3) the corresponding predicted mechanical performance data.

[0139] Designers can choose the appropriate structure for actual manufacturing and application scenarios from multiple solutions that meet the target performance requirements but have different topologies, based on actual application needs.

[0140] This embodiment utilizes a reinforcement learning generation strategy to automatically generate topologies and combines it with a mechanical performance proxy model for rapid evaluation and screening. This achieves automatic mapping from target mechanical performance requirements to heterogeneous lattice structure topology schemes, ensuring that structural performance meets requirements while increasing the diversity of design schemes.

[0141] Example 2 To verify the effectiveness of the heterogeneous lattice structure generative reverse design method proposed in this application, this embodiment conducts experimental verification based on the digital design space, mechanical performance proxy model, and reinforcement learning generation strategy established in Embodiment 1. In this embodiment, the generation of all candidate codes adopts the probabilistic sampling strategy defined in section S106, that is, random sampling according to the action value distribution. Take 1.2.

[0142] First, a lattice structure guided by single-condition, single-direction performance targets is generated in reverse. This addresses the design requirement of precisely matching single-direction performance under specific operating conditions. Target performance values, such as compressive load capacity and energy absorption value, are embedded into the reward function of Example 1. By adjusting the weighting coefficients, the reward gain in the neighborhood of the target performance is enhanced, driving the agent to generate the optimal encoding that satisfies the precise target. The specific implementation method is as follows: taking the X-direction compression condition as an example, the target bearing capacity is set. N, target energy absorption N·mm. Set in the reward function. Disabling anisotropic constraints means Load the trained policy network and adopt the aforementioned probability sampling strategy, i.e., according to... One hundred candidate codes were randomly sampled and generated. A surrogate model was used to predict the corresponding performance. The five schemes with the smallest errors were selected as follows: [2,2,2,3,3,3,1,3], [3,3,1,2,3,1,1,2], [3,1,1,3,3,2,2,1], [2,3,3,2,1,1,1,3], and [2,3,2,3,3,1,2,1]. The finite element simulation results of the optimal schemes are shown below. Figure 9 As shown, the actual load-bearing capacity is 50.526 N, and the energy absorption is 149.688 N·mm, with relative errors of 1.05% and 0.21%, respectively. This model is applicable to the reverse generation of lattice structures with single-direction performance targets under specific working conditions.

[0143] Secondly, a lattice structure is generated inversely based on single-condition, multi-directional performance target orientation. This addresses the design requirement of simultaneously satisfying performance indicators along multiple different directions or planes under the same operating condition. Weighting coefficients are assigned to the performance terms in each direction within the reward function, and lower bound hard constraints or smoothing penalty terms can be introduced to drive the agent to generate optimal encoding that coordinates performance requirements across multiple directions.

[0144] The specific implementation method is as follows: Taking compression as an example, performance targets are set in three directions: bearing capacity in the X direction. Bearing capacity in the N and Y directions Energy absorption in the N and Z directions N·mm. The reward function uses a weighted error summation form, typically with weights in the X direction. Y-direction weights To balance the optimization difficulty, a smoothing penalty function is used for energy absorption in the Z direction, typically with weights... The penalty increases significantly when the predicted value is lower than the target value. The pre-trained policy network is loaded, and 200 candidate codes are generated using the aforementioned probability sampling strategy. These codes are then selected based on their performance. The five schemes with the smallest N·mm and the smallest overall error are [3,2,2,2,3,1,3,3], [2,2,2,2,3,3,3,3], [2,2,2,3,1,3,3,3], [1,1,3,3,3,2,1,2], and [3,3,1,3,2,3,2,2]. The optimal encoding is obtained through finite element simulation and the results are as follows. Figure 10 As shown, N, relative error 1.05%, N, relative error 3.26%, The N·mm value meets the lower limit requirement. This mode achieves precise matching of multi-directional mechanical properties under a single working condition.

[0145] Finally, the reverse generation of multi-condition performance objectives and integrated anisotropic constraints is addressed. Design requirements are set for performance index matching involving multiple loading conditions simultaneously, explicit control of the anisotropy ratio, and the combined constraints of both. The performance error term and anisotropy ratio deviation term under each loading condition are integrated into the reward function, and multi-objective collaborative optimization is achieved by adjusting the weights of each term.

[0146] Simultaneously, it requires precise matching of mechanical properties in specific directions or planes under different working conditions. Taking compression, bending, and torsion as examples, the performance target is set as the bearing capacity in the compression X direction. N, bending moment in the XY plane N·mm, torsional torque in the XY plane N·mm. The reward function assigns equal weights to the three performance terms, typically 1. Loading the policy network, 200 candidate codes are generated using the aforementioned probabilistic sampling strategy. Screening yields the five schemes with the smallest overall error: [2,1,1,1,1,1,1,1,2], [3,1,2,1,1,2,2,3], [1,3,2,2,2,3,1,2], [2,2,1,1,3,2,1,1], [2,2,1,3,1,1,2,1]. The optimal coding finite element simulation and results are as follows: Figure 11 As shown, where N, error 3.2%, N·mm, error 2.2%, N·mm, error 2.9%. This task achieved coordinated and precise matching of performance across operating conditions.

[0147] Facing design requirements with target anisotropy ratios under multiple operating conditions. For example, the compression condition requires isotropic bearing capacity in three directions. Furthermore, the average bearing capacity is close to 45 N; the bending condition requires the bending moment in the YZ plane to be more than 30% greater than that in the other two planes. In the reward function, the compression performance weight is set to 5, and the anisotropic constraint weight is... Set the weight to 10; set the bending performance weight to 15; and the anisotropic constraint weight... Let's set it to 5. Using the aforementioned probabilistic sampling strategy, 200 candidate codes are generated. After screening, the five optimal schemes are obtained: [3,3,2,3,2,3,2,3], [1,3,3,3,2,3,2,3], [3,3,2,3,2,1,2,3], [3,2,2,3,2,3,2,3], and [3,3,2,3,2,3,1,3]. The generated typical structure and finite element simulation results are shown below. Figure 12 As shown, the compression anisotropy ratio Average load-bearing capacity: 45.25 N; bending anisotropy ratio: ,satisfy Requirements. This task achieves quantitative control over the degree of anisotropy.

[0148] In summary, this embodiment verifies that the method proposed in this application can establish a mapping relationship between the topology and mechanical properties of heterogeneous lattice structures, and utilizes reinforcement learning to achieve automated structure generation oriented towards target performance constraints. Compared with the traditional design method that relies on manual experience adjustment and repeated finite element verification, it improves the efficiency and diversity of reverse design of heterogeneous lattice structures.

[0149] Based on the same inventive concept, this application also provides a heterogeneous lattice structure generative reverse design apparatus for implementing the heterogeneous lattice structure generative reverse design method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more heterogeneous lattice structure generative reverse design apparatus embodiments provided below can be found in the limitations of the heterogeneous lattice structure generative reverse design method described above, and will not be repeated here.

[0150] In one embodiment, a heterogeneous lattice structure generative reverse design apparatus is provided, the apparatus comprising: The vector establishment module is used to establish a digital design space for heterogeneous lattice structures. It performs discrete encoding on each basic unit cell that makes up the heterogeneous lattice structure and constructs a topological encoding vector based on the spatial position of each basic unit cell in the heterogeneous lattice structure, so that the topological encoding vector and the heterogeneous lattice structure are uniquely corresponded. The sample construction module is used to construct heterogeneous lattice structure samples based on topology encoding vectors, and to obtain the mechanical performance data of each heterogeneous lattice structure sample under multiple loading conditions and multiple loading directions, so as to construct training samples between topology encoding vectors and mechanical performance data. The model building module is used to build multiple mechanical performance proxy models corresponding to different loading conditions and different loading directions based on training samples, so that the mechanical performance proxy models can predict the target mechanical performance of the corresponding heterogeneous lattice structure according to the input topological encoding vector. The process construction module is used to construct the generation process of the heterogeneous lattice structure as a sequential decision-making process that determines the type of each spatial cell according to a preset cell filling order. The topological encoding vector of the partially filled cell and the current position to be filled are used as the state of the reinforcement learning agent, the cell type corresponding to the current position to be filled is used as the action, and the mechanical performance proxy model is used as the performance evaluator in the reinforcement learning environment. The preset cell filling order is one of a variety of candidate filling orders, and different filling orders correspond to different spatial position traversal paths. The strategy acquisition module is used to construct reward information based on the deviation between the target mechanical performance predicted by the mechanical performance proxy model and the preset target mechanical performance. The reward information is used to train the reinforcement learning agent to obtain the heterogeneous lattice structure generation strategy. The performance prediction module is used to generate multiple candidate topology encoding vectors using the trained heterogeneous lattice structure generation strategy, predict the target mechanical performance of the heterogeneous lattice structure corresponding to each candidate topology encoding vector through a mechanical performance proxy model, and select multiple heterogeneous lattice structure design schemes that meet the target mechanical performance requirements and have different topologies.

[0151] This application also provides an electronic device. In some embodiments, the electronic device includes an input unit, a memory, a processor, and an output unit. The memory stores program instructions that can be executed by the processor. The processor can execute the heterogeneous lattice structure generative reverse design method and / or technical solution based on the foregoing embodiments by calling the program instructions. This electronic device can be a mobile terminal device such as a mobile phone or a computer.

[0152] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program that executes a generative reverse design method for heterogeneous lattice structures. For example, computer program instructions, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions that invoke the methods of this application may be stored in a fixed or removable storage medium, and / or transmitted via data streams in broadcast or other signal carrying media, and / or stored in a storage medium that operates according to the program instructions.

[0153] 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.

[0154] 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.

[0155] 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 generative reverse design method for heterogeneous lattice structures, characterized in that, The method includes: A digital design space for a heterogeneous lattice structure is established, each basic unit cell that makes up the heterogeneous lattice structure is discretely encoded, and a topological encoding vector is constructed based on the spatial position of each basic unit cell in the heterogeneous lattice structure, so that the topological encoding vector establishes a unique correspondence with the heterogeneous lattice structure. Based on the topology encoding vector, heterogeneous lattice structure samples are constructed, and mechanical performance data of each heterogeneous lattice structure sample under multiple loading conditions and multiple loading directions are obtained to construct training samples between the topology encoding vector and the mechanical performance data. Based on the training samples, multiple mechanical performance proxy models are established, each corresponding to different loading conditions and different loading directions, so that the mechanical performance proxy models can predict the target mechanical performance of the corresponding heterogeneous lattice structure according to the input topological encoding vector. The generation process of the heterogeneous lattice structure is constructed as a sequential decision-making process that determines the type of each spatial cell according to a preset cell filling order. The topological encoding vector of the partially filled cell and the current position to be filled are used as the state of the reinforcement learning agent, the cell type corresponding to the current position to be filled is used as the action, and the mechanical performance proxy model is used as the performance evaluator in the reinforcement learning environment. The preset cell filling order is one of a variety of candidate filling orders, and different filling orders correspond to different spatial position traversal paths. Reward information is constructed based on the deviation between the target mechanical performance predicted by the mechanical performance proxy model and the preset target mechanical performance. The reward information is used to train the reinforcement learning agent to obtain a heterogeneous lattice structure generation strategy. Multiple candidate topology encoding vectors are generated using the trained heterogeneous lattice structure generation strategy. The target mechanical properties of the heterogeneous lattice structure corresponding to each candidate topology encoding vector are predicted by the mechanical performance proxy model. Multiple heterogeneous lattice structure design schemes that meet the target mechanical performance requirements and have different topologies are selected.

2. The heterogeneous lattice structure generative reverse design method as described in claim 1, characterized in that, Establishing a digital design space for heterogeneous lattice structures includes: Multiple basic unit cells with different mechanical behavior characteristics are selected as the constituent units of the heterogeneous lattice structure, and each basic unit cell is assigned a unique discrete unit cell type code. Establish the spatial layout corresponding to each basic unit cell in the heterogeneous lattice structure, and map each spatial position to the position of each element in the topological coding vector according to a fixed spatial coding order, so that each element position in the coding vector corresponds one-to-one with each spatial position. Record the discrete unit cell type codes corresponding to each spatial location in the fixed spatial coding order to form a topological coding vector; Each topological coding vector uniquely corresponds to a heterogeneous lattice structure, and each heterogeneous lattice structure can be mapped to a unique topological coding vector. All topological coding vectors together constitute the digital design space of the heterogeneous lattice structure.

3. The heterogeneous lattice structure generative reverse design method as described in claim 2, characterized in that, Constructing the training samples includes: The parameterized finite element model of the corresponding heterogeneous lattice structure is automatically generated based on the topology encoding vector; Finite element analysis models for compression, torsion, and bending conditions were established respectively, and mechanical simulations were performed along multiple loading directions or multiple loading planes respectively. Extract the mechanical response curves corresponding to each finite element analysis model, and calculate the compressive bearing capacity, torsional torque, bending moment and anisotropy ratio corresponding to each loading condition and loading direction based on the mechanical response curves to form mechanical performance data corresponding to the heterogeneous lattice structure. The topological encoding vector is associated with the corresponding mechanical performance data to construct training samples for training the mechanical performance proxy model.

4. The heterogeneous lattice structure generative reverse design method as described in claim 3, characterized in that, Establishing the mechanical property proxy model includes: Feature engineering is performed on the topological coding vector to convert the discrete topological code into a continuous feature vector. The continuous feature vector includes at least basic normalized features, statistical features, spatial correlation features, symmetry features, unit cell type distribution features, and global aggregation features. The continuous feature vector is then standardized. Deep feedforward neural networks are constructed for different loading conditions and different loading directions. Each deep feedforward neural network includes an input layer, a shared feature extraction layer, multiple task branch layers and an output layer. The shared feature extraction layer is used to extract common features of the topology, and each task branch layer is used to predict the load-bearing capacity, torque or energy absorption performance under the corresponding working conditions and loading directions. The corresponding training samples are used to train each deep feedforward neural network. When each deep feedforward neural network meets the preset prediction accuracy, the network parameters are fixed to form the multiple mechanical performance proxy models. When the deep feedforward neural network simultaneously predicts both load-bearing capacity and energy absorption capacity, the loss function is: The mean square error between the predicted and actual carrying capacity values. The mean square error between the predicted and actual values ​​of energy absorption capacity. and These are weighting coefficients, generally = =0.

5.

5. The heterogeneous lattice structure generative reverse design method as described in claim 4, characterized in that, Obtaining strategies for generating heterogeneous lattice structures includes: A finite-step sequential decision-making process is established based on the preset cell filling order. The topological encoding vector of the current part being filled and the current position to be filled are used together as the state of the reinforcement learning agent, and the basic cell type corresponding to the current position to be filled is used as the action. After completing all unit cell filling, the generated complete topology encoding vector is input into the corresponding mechanical performance proxy model to obtain the predicted mechanical performance under the target working condition and target loading direction. The terminal reward is constructed based on the deviation between the predicted mechanical performance and the target mechanical performance. Anisotropic constraints are calculated based on the predicted mechanical performance corresponding to multiple loading conditions and multiple loading directions. The terminal reward and anisotropic constraints are used together to determine the reward value of the reinforcement learning environment. The reward function is as follows: ； in For the round reward function, , For the first One performance prediction value, For the target value, This is the weighting coefficient, usually set to 1. The total number of target performance metrics of interest; The anisotropy ratio is calculated based on multi-directional prediction values. The target is the anisotropy ratio; To prevent division by zero for small constants; Weights for anisotropic constraints; A reinforcement learning agent based on the Dueling DQN architecture is trained interactively with the reinforcement learning environment, and the action value network parameters are updated by combining the experience replay mechanism until a heterogeneous lattice structure generation strategy that satisfies the target mechanical performance constraints is obtained.

6. The heterogeneous lattice structure generative reverse design method as described in claim 5, characterized in that, Multiple heterogeneous lattice structure design schemes were generated, including: Load the trained heterogeneous lattice structure generation strategy and output the action value corresponding to each candidate action based on the current state; The action probability is determined based on the action value corresponding to each candidate action, and the basic cell type corresponding to each spatial location is randomly sampled according to the action probability to generate multiple topological coding vectors. The mechanical performance of the target mechanical properties of the heterogeneous lattice structure corresponding to each topological coding vector is predicted using the aforementioned mechanical performance proxy model. Based on the prediction results, the topology coding vectors that meet the target mechanical performance constraints are selected, and the topology repeatability of the topology coding vectors that meet the target mechanical performance constraints is checked. After deleting duplicate topology coding vectors, multiple heterogeneous lattice structure design schemes that meet the same target mechanical performance constraints but have different topologies are output.

7. The heterogeneous lattice structure generative reverse design method as described in claim 6, characterized in that, The preset unit cell filling order includes multiple filling orders determined by various spatial traversal methods; During the reinforcement learning strategy training process, the various filling orders are used for training to verify the robustness of the heterogeneous lattice structure generation strategy to different filling orders, and the optimal filling order or the strategy knowledge of multiple orders is determined based on the training results under each filling order.

8. A heterogeneous lattice structure generative reverse design device, characterized in that, The device includes: The vector establishment module is used to establish a digital design space for the heterogeneous lattice structure, to discretely encode each basic unit cell that makes up the heterogeneous lattice structure, and to construct a topological encoding vector based on the spatial position of each basic unit cell in the heterogeneous lattice structure, so that the topological encoding vector establishes a unique correspondence with the heterogeneous lattice structure. The sample construction module is used to construct heterogeneous lattice structure samples based on the topology encoding vector, and to obtain the mechanical performance data of each heterogeneous lattice structure sample under multiple loading conditions and multiple loading directions, so as to construct training samples between the topology encoding vector and the mechanical performance data. The model building module is used to build multiple mechanical performance proxy models corresponding to different loading conditions and different loading directions based on the training samples, so that the mechanical performance proxy models can predict the target mechanical performance of the corresponding heterogeneous lattice structure according to the input topological encoding vector. The process construction module is used to construct the generation process of the heterogeneous lattice structure as a sequential decision-making process that determines the type of each spatial cell according to a preset cell filling order. The topological encoding vector of the partially filled cell and the current position to be filled are used as the state of the reinforcement learning agent, the cell type corresponding to the current position to be filled is used as the action, and the mechanical performance proxy model is used as the performance evaluator in the reinforcement learning environment. The preset cell filling order is one of a variety of candidate filling orders, and different filling orders correspond to different spatial position traversal paths. The strategy acquisition module is used to construct reward information based on the deviation between the target mechanical performance predicted by the mechanical performance proxy model and the preset target mechanical performance, and to train the reinforcement learning agent with the reward information to obtain the heterogeneous lattice structure generation strategy. The performance prediction module is used to generate multiple candidate topology encoding vectors using the trained heterogeneous lattice structure generation strategy, predict the target mechanical performance of the heterogeneous lattice structure corresponding to each candidate topology encoding vector through the mechanical performance proxy model, and select multiple heterogeneous lattice structure design schemes that meet the target mechanical performance requirements and have different topologies.

9. 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 heterogeneous lattice structure generative reverse design method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the heterogeneous lattice structure generative reverse design method according to any one of claims 1 to 7.

Images