Hybrid architecture mechanical property prediction method and system for additive manufacturing lattice structure

Abstract

The invention relates to a hybrid architecture mechanical property prediction method and system for an additive manufacturing lattice structure, and the method comprises the steps: converting a lattice structure node three-dimensional coordinate into a normalized distance from a lattice center, introducing a self-attention mechanism, and constructing a feature prediction model in combination with residual connection and layer normalization; constructing a generative adversarial network comprising a generator and a discriminator; a strategy network based on prediction error rewards is constructed, and a depth deterministic strategy gradient algorithm is adopted for optimization; jointly training the models, and constructing a performance prediction feedback model; finally, the node coordinates of the target structure serve as input, and mechanical property prediction output is achieved. According to the method, the generative adversarial neural network, reinforcement learning and an attention mechanism are mutually combined to generate a hybrid architecture, the problems of gradient disappearance and mode collapse of a traditional GAN can be effectively relieved, the model can accurately recognize implicit association between node positions and mechanical properties, and then the mechanical properties of target lattice structure parameters are rapidly predicted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lattice structure performance prediction, and more particularly to a hybrid architecture mechanical performance prediction method and system for additive manufacturing lattice structures. BACKGROUND

[0002] In recent years, additive manufacturing technology with free manufacturing capability has gradually matured, making the manufacturing of complex lightweight structures such as lattice structures possible. Compared with traditional processing methods, additive manufacturing breaks the original design restrictions, enabling the production of complex parts while reducing production cycle and processing cost. The high strength, lightweight and excellent thermal control characteristics of lattice structures and the maturity of additive manufacturing technology make it a new idea for lightweight structure design. The combination of three-dimensional lattice structures and additive manufacturing technology has become a new method for designing and manufacturing high-performance lightweight multifunctional structures.

[0003] Lattice structure is a topologically ordered three-dimensional structure composed of multiple repeating cells arranged periodically in space, with highly adjustable geometric configuration and great design freedom. However, this advantage also brings unprecedented engineering challenges: the number of theoretically feasible lattice structure configurations is almost infinite, and the computational cost of performance prediction is extremely high, which becomes a key bottleneck restricting its large-scale application.

[0004] Currently, the performance prediction of lattice structures mainly relies on the finite element simulation (FEM) method, which models, loads and analyzes each structure to obtain mechanical indicators such as elastic modulus and compressive strength. Although this method has high accuracy, each configuration needs to be modeled and simulated separately, which is difficult to meet the needs of high-throughput design and evaluation. The existing simulation results lack generalization ability and cannot be directly used for performance estimation of new configurations. Moreover, the mapping relationship between lattice structure geometric parameters and mechanical properties is often highly nonlinear and coupled, making it difficult to accurately predict using traditional linear analysis and statistical regression methods. SUMMARY

[0005] The present application provides a hybrid architecture mechanical performance prediction method and system for additive manufacturing lattice structures, which combines generative adversarial neural networks (GAN), reinforcement learning and attention mechanisms to generate a hybrid architecture. This can effectively alleviate the gradient vanishing and mode collapse problems of traditional GANs, enabling the model to accurately identify the implicit relationship between node positions and mechanical properties, and then quickly predict the mechanical properties of target lattice structure parameters.

[0006] In a first aspect, the application provides a hybrid architecture mechanical property prediction method for additive manufacturing lattice structures, the method comprising: converting node three-dimensional coordinates of a plurality of lattice structures constructed in a random manner into normalized distances from the center of the lattice, then introducing a self-attention mechanism and combining a residual connection and a layer normalization mechanism to construct a feature prediction model; constructing a generative adversarial network model comprising a generator and a discriminator, wherein the generator takes a random noise vector as input and generates latent lattice structure parameters, and the discriminator utilizes a channel attention mechanism and combines an adversarial loss and a mean square error loss to optimize the feature recognition capability; using the prediction error as a reward function and adopting a deep deterministic policy gradient algorithm to construct a policy network; jointly training the generative adversarial network model, the feature prediction model and the policy network to construct a lattice structure mechanical property prediction feedback model; inputting node three-dimensional coordinates of a target lattice structure into the lattice structure mechanical property prediction feedback model to predict and output the mechanical property of the target lattice structure.

[0007] In an optional implementation of the first aspect, when constructing the feature prediction model, the method comprises: randomly generating node three-dimensional coordinates of a preset number of lattice structures and converting the node three-dimensional coordinates into normalized distances from the center of the lattice to highlight the position sensitivity in the horizontal direction; introducing an interaction term and a high-order polynomial term to enhance the nonlinearity representation; introducing a self-attention mechanism for dynamically assigning different position feature weights: wherein Q is a query matrix generated by linear mapping of input features through a weight matrix, K is a keyword matrix generated by linear mapping of input features to represent attention position identification information, V is a value matrix generated by linear mapping of input features to represent original information of each position, QK T is a similarity score, d k is a scaling factor; then generating an attention-weighted feature vector; inputting the attention-weighted feature vector into a fully connected network and combining a residual connection and a layer normalization mechanism to construct a feature prediction model for predicting mechanical properties.

[0008] In an optional implementation of the first aspect, when introducing the self-attention mechanism for dynamically assigning different position feature weights, the method comprises: calculating the sensitivity decay weight of each node three-dimensional coordinate position according to the normalized distance converted from the node three-dimensional coordinates: wherein w i is the sensitivity decay weight of the i-th node, d i is the normalized distance of the i-th node, and a is a sensitivity decay factor for adjusting the decay intensity; constructing a symmetric position sensitivity matrix: w j is the sensitivity decay weight of the j-th node; applying the symmetric position sensitivity matrix to the self-attention mechanism to realize a position sensitivity weighted attention mechanism:

[0009] where is the Hadamard product, used to dynamically adjust each attention score according to the sensitivity decay weight of the two nodes.

[0010] In an optional implementation of the first aspect, in constructing the generative adversarial network model comprising the generator and the discriminator, the method comprises: constructing the generator taking a random noise vector as input and generating a latent point lattice structure parameter, and introducing a spatial attention mechanism to preferentially focus on the node distribution in the horizontal direction in the generation process; constructing the discriminator adopting a channel attention mechanism dynamically adjusting the weight of different feature channels and combining an adversarial loss and a mean square error loss to optimize the feature recognition capability; and constructing the generative adversarial network model according to the generator and the discriminator.

[0011] In an optional implementation of the first aspect, in constructing the strategy network, the method comprises: constructing the strategy network dynamically adjusting the attention weight distribution of the feature prediction model according to the difference between the predicted performance and the target performance by using a deep deterministic policy gradient algorithm; and defining a reward function of the strategy network as the negative value of the prediction error:

[0012] R = -(|E pred -E true | + |s pred -s true |), where R is the reward function, E pred is the predicted elastic modulus, E true is the true elastic modulus, s pred is the predicted compressive strength, and s true is the true compressive strength.

[0013] In an optional implementation of the first aspect, in training the generative adversarial network model, the feature prediction model, and the strategy network, the method comprises: generating a point lattice structure parameter by using the generative adversarial network model and predicting the mechanical performance of the candidate point lattice structure by using the feature prediction model to generate a prediction result; screening samples with a preset mechanical performance greater than a preset threshold according to the prediction result; updating the strategy network parameters to optimize the generation and attention allocation strategy in the next round; training the feature prediction model separately using a preset training data set and fixing the feature prediction model parameters; jointly optimizing the generator and the discriminator of the generative adversarial network model and introducing a gradient penalty term to alleviate the training instability problem; fine-tuning the strategy network parameters on the basis of freezing the generator and the feature prediction model; and repeating the training process until a preset iteration number or a preset index is reached, and then constructing the point lattice structure mechanical performance prediction feedback model by jointly training the completed generative adversarial network model, the feature prediction model, and the strategy network.

[0014] In an optional implementation of the first aspect, the method for generating the preset training data set comprises: constructing a three-dimensional model of a random dot lattice structure according to three-dimensional coordinates of a preset number of randomly generated unit nodes; calculating a stress-strain curve of the random dot lattice structure by using finite element simulation; and extracting an elastic modulus and a compressive strength according to the stress-strain curve to form the training data set.

[0015] In an optional implementation of the first aspect, when optimizing the generator of the generative adversarial network model, the method comprises: calculating Euclidean distances between all nodes and taking variances of the Euclidean distances as a node distribution uniformity index:

[0016] wherein N is a total number of nodes, is an average distance between all nodes, D ij is a Euclidean distance between an i th node and a j th node, D ij =‖x i -x j ‖2, x i is a three-dimensional coordinate of the i th node, x j is a three-dimensional coordinate of the j th node; rearranging the dot lattice structure under a preset geometric transformation function and performing point set matching with the original dot lattice structure to calculate a configuration symmetry index:

[0017] wherein T k is a geometric transformation function, S k is an average minimum matching error of the dot lattice structure under a k th transformation with the original dot lattice structure, and K is a number of geometric transformations; adding a weighted structure prior constraint term in a loss function of the generator to realize guided optimization of the uniformity and symmetry of the dot lattice structure parameters: wherein, is a total loss function of the generator, L adv is an adversarial loss function, λ U and λ S are hyperparameters for adjusting weights of the structure prior constraint term in the total loss.

[0018] In an optional implementation of the first aspect, when optimizing the discriminator of the generative adversarial network model, the method comprises: inputting dot lattice structure parameters generated by the generator and real dot lattice structure parameters into the discriminator; generating a discrimination output by the discriminator; calculating a total loss function of the discriminator by using a weighted combination of an adversarial loss function, a mean square error loss function and a structural similarity loss function; and updating the discriminator parameters by using a gradient descent method or an Adam optimization method according to the total loss function of the discriminator.

[0019] In a second aspect, the application provides a hybrid architecture mechanical property prediction system for additive manufacturing lattice structure using the method, comprising: a feature extraction module for converting the node three-dimensional coordinates of a plurality of lattice structures constructed randomly into normalized distances from the lattice center, then introducing a self-attention mechanism and combining a residual connection with a layer normalization mechanism to construct a feature prediction model; a GAN construction module for constructing a generative adversarial network model comprising a generator and a discriminator, wherein the generator takes a random noise vector as input and generates latent lattice structure parameters, and the discriminator utilizes a channel attention mechanism and combines an adversarial loss with a mean square error loss to optimize the feature recognition ability; a dynamic feedback module for using the prediction error as a reward function and constructing a policy network using a deep deterministic policy gradient algorithm; a joint modeling module for jointly training the generative adversarial network model, the feature prediction model and the policy network to construct a lattice structure mechanical property prediction feedback model; and a performance prediction module for inputting the node three-dimensional coordinates of a target lattice structure into the lattice structure mechanical property prediction feedback model to predict the mechanical properties of the target lattice structure.

[0020] It should be understood that the general description above and the following detailed description are only exemplary and do not limit the application. BRIEF DESCRIPTION OF DRAWINGS

[0021] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate one or more embodiments of the application and, together with the description, serve to explain the principles of the application and to enable a person skilled in the relevant art to make and use the application.

[0022] Figure 1 is a schematic diagram of the pain points according to the prior art optimization design method; wherein (a) is a traditional trial-and-error optimization method, and (b) is a missing rod and powder accumulation in actual production of lattice structure.

[0023] Figure 2 is a schematic diagram of an exemplary stress-strain curve of a random structure according to some embodiments of the application.

[0024] Figure 3 is a schematic diagram of the analysis of an exemplary modulus and compressive strength training data set and the selection of a machine learning algorithm according to some embodiments of the application.

[0025] Figure 4 is a schematic diagram of the structure of an exemplary GAN network model according to some embodiments of the application.

[0026] Figure 5 is a schematic diagram of the flow of an exemplary hybrid architecture mechanical property prediction method for additive manufacturing lattice structure according to some embodiments of the application.

[0027] Figure 6 is an exemplary module connection diagram of a hybrid architecture mechanical property prediction system for additive manufacturing lattice structures according to some embodiments of the present application.

[0028] Figure 7 is an exemplary diagram of the impact of machine learning hyperparameters on model performance according to some embodiments of the present application; where (a) is the elastic modulus prediction model hyperparameter tuning, (b) is the compressive strength model hyperparameter tuning, (c) is the elastic modulus prediction model learning curve, and (d) is the compressive strength prediction model learning curve. Machine learning prediction is excellent.

[0029] Figure 8 is an exemplary diagram of the elastic modulus predicted value versus the true value according to some embodiments of the present application.

[0030] Figure 9 is an exemplary diagram of the compressive strength predicted value versus the true value according to some embodiments of the present application.

[0031] Figure 10 is an exemplary diagram of the relationship between elastic modulus and node position according to some embodiments of the present application.

[0032] Figure 11 is an exemplary diagram of the relationship between compressive strength and node position according to some embodiments of the present application.

[0033] Figure 12 is an exemplary connection diagram of a terminal device according to some embodiments of the present application. DETAILED DESCRIPTION

[0034] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations can be implemented in any

[0035] In recent years, the additive manufacturing technology with free manufacturing capability has gradually matured, making it possible to manufacture complex lightweight structures such as lattice structures. Compared with traditional processing methods, additive manufacturing breaks the original design restrictions, enabling the production of complex parts while reducing production cycle and processing costs. Laser powder bed fusion (LPBF) is one of the most promising methods for manufacturing complex metal parts with additive manufacturing technology. The density of parts manufactured by LPBF process can reach 99.9%, and the mechanical properties are closer to those of parts manufactured by traditional methods. The outstanding advantage of LPBF process for lattice structure forming is the release of the design freedom of lattice structure, which can design and manufacture lattice structures with multiple functions such as bearing, energy absorption, vibration reduction, and damping. It is an important embodiment of current material-structure-process integrated design and manufacturing.

[0036] Lattice structure is a topologically ordered three-dimensional structure composed of multiple repeating cells arranged periodically in space. The high strength, lightweight, and excellent thermal control characteristics of lattice structure, as well as the maturity of additive manufacturing technology, make it a new idea for lightweight structure design. The combination of three-dimensional lattice structure and additive manufacturing technology has become a new method for designing and manufacturing high-performance lightweight multifunctional structures. Metal lattice structure based on additive manufacturing technology realizes lightweight preparation and performance improvement of the structure, greatly shortening the production cycle of complex structures. The excellent performance of metal lattice structure and the rapid preparation of additive manufacturing technology have made it a research hotspot in the field of aerospace and other fields.

[0037] The materials used for preparing metal lattice structures are mainly titanium alloy, aluminum alloy and stainless steel, which are limited by the materials available for additive manufacturing. Among them, Ti-6Al-4V alloy is one of the most widely used alloys for preparing high lightweight lattice structures due to its excellent overall performance. The mechanical properties and forming quality of metal lattice structures vary with different cell arrangement, cell lattice size, rod diameter, relative density and materials. The mechanical response is complex and non-uniform, and there is no performance test standard that can be completely applied to lattice structures. GB / T31930-2015

[21] “Metallic materials ductility test Porous and cellular metal compression test method”(same as ISO13314:2011“Mechanical testing of metals-Ductility testing-Compression test for porous and cellular metals”) provides a test method for measuring the compression performance of porous structures, but this standard is only effective for structures with porosity greater than 50%. The relative density of lattice structure design is usually between 20% and 80%, and this standard can only cover part of the lattice structure. At the same time, the design parameters of lattice structure are numerous, and the existing mechanical performance test models and test methods are different, making the performance parameters of lattice structures in different studies not comparable. Due to the imperfect test standards for lattice structures and the limitations of tool clamps, the current research on the mechanical properties of lattice structures mainly focuses on compressive strength, yield strength, elastic modulus and fatigue performance. The related research on the mechanical properties of metal lattice structures shows that with the increase of relative density and aspect ratio, the compressive capacity of lattice structure increases, the elastic modulus becomes larger, the relative stiffness increases, and the bending and shear effect is more obvious. When the relative density is greater than 0.1, the performance is no longer linearly related to the relative density, and the classical linear prediction model is no longer applicable, and a new mechanical model needs to be established. The main factors affecting the mechanical properties currently found are relative density, node effect, bending shear effect and surface powder, among which the main factors affecting the size of the elastic modulus of the lattice structure are the equivalent Young's modulus of the matrix material, the ratio of rod diameter to rod length, and the main factors affecting the size of the Poisson's ratio of the lattice structure are the length of the cell and the aspect ratio.

[0038] Lattice structures will play an important role in the design of lightweight and high-performance equipment in the future, have a wide range of application scenarios, and will help improve the performance of aerospace equipment and other fields. However, due to the uncertainty of the preparation process and the complexity of the structure design, it is difficult and slow to predict and optimize the performance, which restricts the iteration speed of related application equipment. In the traditional experimental trial-and-error optimization, the designed lattice structure needs to be produced, the performance needs to be tested, and the structure needs to be optimized. The mechanical property test is a destructive experiment, and a large number of samples are required. After tens or even nearly a hundred iterations of optimization, the structure parameters that meet the performance requirements can be obtained. The optimization method combined with numerical simulation has no great change compared with the traditional experimental trial-and-error process. This method saves the process of sample preparation and mechanical property testing to a certain extent, but increases the threshold requirement of optimization design. The efficiency of numerical simulation is low, and the time cost of the optimization process cannot be reduced. At the same time, since the numerical simulation process is usually based on the ideal geometric conditions of the lattice structure model, it cannot consider the performance impact of the defects inevitably caused by the additive manufacturing process, as shown in Figure 1 Figure 1 The pain points of the prior optimization design method are shown in the schematic diagram, wherein Figure 1 (a) is a traditional trial-and-error optimization method, Figure 1 (b) is the missing of the rod and the powder accumulation in the actual production of the lattice structure, which leads to a large error between the simulation results and the actual performance, and the significant fluctuation of the prediction results.

[0039] Therefore, in the face of almost infinite theoretical design configurations of unit cells, it is expensive and impractical to use finite element analysis to test the mechanical behavior of the lattice structure. Using machine learning to accurately predict the performance of more structures is an important means to quickly obtain the performance of unknown structures. The present application aims to quickly predict the mechanical performance of metal component lattice structures, uses machine learning methods, uses data modeling and simulation to obtain training data, and builds an ensemble learning prediction model to predict and verify the compression performance of a single-node lattice structure with a relative density of 20%.

[0040] Specifically, in order to quickly accumulate training data, the present application uses a method of randomly generating unit cells and automatically performing modeling, simulation and post-processing. The computing device first randomly generates multiple groups of three-dimensional coordinates containing three numbers between 0 and 1, then uses finite element simulation to establish a three-dimensional model, and then performs simulation calculation by means of ABAQUS, and exports the stress-strain curve as shown in Figure 2 , and finally calculates the elastic modulus and compressive strength mechanical properties as training data as shown in Figure 3 .

[0041] ​To select a suitable machine learning prediction algorithm, the data set is first analyzed. Determine whether there is a linear relationship between the three-dimensional coordinate factors of the node and the mechanical properties. Sum of Squares (SS) reflects the total variability or deviation of the data. By calculating the contribution of a factor or error term to the total variability, the significance of different factors can be evaluated. The formula is: Where y i is the i-th observation, is the average of all observations.

[0042] In analysis of variance, F-Value is used to test whether a factor has a significant effect on the result. It measures the ratio of between-group variation to within-group variation. P-Value is used to measure the significance of a factor, representing the probability of observing the current result under the assumption of no significance (null hypothesis). When P-Value is less than 0.05, the null hypothesis is rejected, indicating that the factor has a significant effect on the result.

[0043] The results of the variance analysis of the node coordinate factors and the elastic modulus are shown in the following table:

[0044] Table: Variance analysis of node coordinate factors and elastic modulus

[0045]

[0046] The results of the variance analysis of the node coordinate factors and the compressive strength are shown in the following table:

[0047] Table: Variance analysis of node coordinate factors and compressive strength

[0048]

[0049] Among them, the three-dimensional coordinates of the node and their interactions have no significant effect on the elastic modulus and compressive strength. All factors do not reach the significant level (P-Value>0.05), and there is a large amount of variation that has not been explained, which may be due to other potential influencing factors or complex nonlinear relationships.

[0050] The above results show that when selecting a machine learning prediction algorithm for the mechanical properties of lattice structures, models that can handle complex nonlinear relationships and high-order interactions should be preferred. Although the variance analysis results show that the three-dimensional coordinates of the node and their interactions have no significant effect on the elastic modulus and compressive strength (P-Value>0.05), this does not mean that these features are worthless in prediction. Traditional linear statistical methods are limited by global assumptions and are difficult to capture local nonlinear effects and high-order feature coupling relationships. Machine learning models can extract potential mechanical response patterns from seemingly unrelated geometric parameters through adaptive feature engineering and nonlinear mapping mechanisms.

[0051] The attention mechanism model is good at extracting high-order features of data, and the complex model prediction regression has strong performance. At the same time, due to the continuous acquisition of a large number of parameters of the model, the simple manual marking of the supervised model is difficult to fully meet the needs of the present application, and the acquisition of a large amount of labeled data is expensive. In data generation, strategy learning and other learning tasks, the acquisition of these labeled data has become a difficult problem. The existing mature decision model GAN network solves this problem well. The GAN neural network is composed of a discriminator and a generator. The generator G receives a random variable z to generate a fake sample data. The purpose of the discriminator is to make the generated sample as real as possible, as shown in reference Figure 4 The two games improve the performance of the model. However, the actual GAN is not mature, and there are many problems, among which the gradient disappearance and mode collapse problems seriously limit the application of the model.

[0052] Therefore, in order to solve the above problems, the present application designs a hybrid architecture of generative adversarial neural network (GAN) combined with reinforcement learning and attention mechanism. The generator generates a virtual dot array configuration consistent with the real data distribution through adversarial training, and combines a new information quantity criterion to filter high-value samples. The discriminator dynamically optimizes the feature weight through reinforcement learning to enhance the representation ability of the nonlinear relationship. The hybrid architecture of the present application can effectively alleviate the gradient disappearance and mode collapse problems of the traditional GAN, so that the model can accurately identify the implicit correlation between the node position and the mechanical properties, for example, the closer the node is to the lattice center (the coordinate approaches 0.5), the higher the elastic modulus and compressive strength will be. Such local nonlinear laws are not revealed by the analysis of variance.

[0053] Reference Figure 5 As shown in the figure, Figure 5 The present application designs a hybrid architecture mechanical property prediction method for additive manufacturing dot array structure 1. In order to realize efficient prediction and feedback of the mechanical properties of the dot array structure, a deep learning architecture based on attention mechanism (Attention Mechanism) is combined with the dynamic optimization strategy of generative adversarial network (GAN) and reinforcement learning to construct a multi-level and high-precision prediction and feedback method. Specifically, the method comprises:

[0054] S1: Convert the three-dimensional coordinates of the nodes of a plurality of dot array structures constructed at random into normalized distances from the lattice center, and then introduce a self-attention mechanism and combine residual connection and layer normalization mechanism to construct a feature prediction model.

[0055] Specifically, when constructing the feature prediction model in S1, the method comprises:

[0056] S11: Randomly generate the node three-dimensional coordinates (x, y, z) of a preset number of lattice structures and convert the node three-dimensional coordinates into normalized distances from the lattice center to highlight the position sensitivity in the horizontal direction.

[0057] wherein, in generating the node three-dimensional coordinates, the designer generates them according to the set lattice structure type (such as structure pattern, lattice unit size, periodic boundary, generation scale) by a regular method (such as a lattice point formula); the predicted number is the number of composite lattice structures designed by the designer according to the actual demand and cost. In converting the node three-dimensional coordinates into normalized distances from the lattice center, let the three-dimensional coordinates of the i-th node be (x i ,y i ,z i ), the center of gravity of all node coordinates be (x c ,y c ,z c ), and the normalized distance of the node be defined as:

[0058] wherein d max is the maximum Euclidean distance between all nodes and the center of gravity, which is used for normalization, and the center of gravity is the lattice center of gravity.

[0059] S12: Introduce interaction terms and high-order polynomial terms to improve the expression ability of the model for the nonlinear relationship between complex geometric distribution and mechanical response.

[0060] S13: Introduce a dynamic allocation of different position feature weights self-attention mechanism (Multi-Head Self-Attention):

[0061] wherein Q is a query matrix generated by linear mapping of the input features through a weight matrix, K is a keyword matrix generated by linear mapping of the input features to represent the attention position identification information, V is a value matrix generated by linear mapping of the input features to represent the original information of each position, QK T is the similarity score, d k is the scaling factor; and then generate an attention weighted feature vector.

[0062] wherein the specific process can be to stack the feature vectors of all nodes into a matrix: F = [f1, f2, …, f N ] T , then construct Q = FW Q , K = FW K , and V = FW V , wherein W Q , W K , and W V are learnable weights.

[0063] S14: The attention-weighted feature vector is input into the fully connected network and combined with the residual connection (ResidualConnection) and layer normalization (LayerNorm) mechanism to construct a feature prediction model for predicting mechanical properties.

[0064] Among them, an efficient nonlinear mapping model is constructed using attention-weighted features to predict the mechanical properties of nodes or overall structures; in the residual connection structure, the output of each layer is added to the input of the previous layer to alleviate the gradient vanishing problem in deep network training; in layer normalization, normalization is performed on the input of each layer so that the features are evenly distributed in different dimensions; the predicted mechanical properties of this application are preferably elastic modulus and compressive strength, that is, the elastic modulus mainly measures the deformation ability of the material in the elastic stage, reflecting the rigidity of the material; and the compressive strength measures the maximum bearing capacity of the material under compressive conditions, reflecting the compressive performance of the material.

[0065] S2: Construct a generative adversarial network model consisting of a generator and a discriminator, where the generator takes a random noise vector as input and generates latent lattice structure parameters, and the discriminator uses a channel attention mechanism and combines adversarial loss with mean square error loss to optimize feature recognition capabilities.

[0066] Specifically, when constructing a generative adversarial network model including a generator and a discriminator in S2, the method includes:

[0067] S21: Construct a generator that takes a random noise vector as input and generates the parameters of the latent lattice structure, and introduces a spatial attention mechanism to prioritize the horizontal node distribution during the generation process.

[0068] The input of the generator is a random noise vector that obeys a Gaussian distribution. The structure of the generator can be set with multiple deconvolution layers, batch normalization layers, and activation functions. In order to enhance the generator's attention to the horizontal arrangement pattern of the structure (such as the transition zone between layers, the lateral support strength, etc.), a spatial attention mechanism is introduced. By performing weighted calculations on the spatial dimensions of the intermediate feature map, priority is given to the distribution pattern of nodes in the horizontal direction. That is, the spatial attention mechanism calculates the significant response of the feature map in the lateral dimension (such as the X direction) to guide the generator to improve the expression ability of the lateral topological features when generating the lattice structure parameters. The spatial attention can take the following form:

[0069] M s =σ(Conv2D([AP c (F); MP c (F)])), where AP c (F); MP c(F) is average pooling and maximum pooling in the channel dimension, Conv2D is a convolution operation to extract spatial relationships, and is a Sigmoid activation function to output normalized weights, and F is a feature map. The output of the generator is a lattice structure topology or geometric parameter, which can be directly used for finite element modeling or structural performance prediction.

[0070] S22: Construct a channel attention mechanism that dynamically adjusts the weights of different feature channels, and combine the adversarial loss and mean square error loss to optimize the feature recognition ability of the discriminator.

[0071] wherein the discriminator receives an input sample and discriminates whether it comes from a real structure sample distribution, extracts structure features through multiple convolution layers and outputs an adversarial prediction score; the input of the discriminator is real and generated lattice structure parameters, and the structure of the discriminator can include multiple convolution layers, pooling layers and activation functions. To enhance the perception of the relationship between the channels of the structure parameters, a channel attention mechanism is introduced to dynamically adjust the weight of each channel in the feature map; wherein the channel attention can take the following form: M c = σ(W2(ReLU(W1([AP s (F) ; MP s (F)])))), wherein AP c (F) ; MP c (F) is average pooling and maximum pooling in the channel dimension, W1 and W2 are two fully connected layers, and is a Sigmoid activation function to output normalized weights; wherein the loss function of the discriminator is the adversarial loss and the mean square error loss, that is:

[0072] l total = l adv + λ·l MSE , wherein l total is the total loss function of the discriminator, l adv is the adversarial loss function to judge true and false samples, l MSE is the mean square error loss function for enhancing the discrimination ability of structure details, and λ is the weight coefficient. The output of the discriminator is a channel weight vector, which is used to adjust the original feature map.

[0073] S23: Construct a generative adversarial network model according to the generator and the discriminator.

[0074] By adopting the technical scheme, the spatial and channel attention mechanism is introduced into the generative adversarial network, so that the generator can pay more attention to key morphological features such as transverse node distribution in the structure construction process, and the sensitivity of the discriminator to local and global structure features is improved, thereby improving the topological configuration simulation and mechanical property mapping capability of the entire network on the additive manufacturing lattice structure, and providing high-quality sample support for data enhancement, topological optimization and subsequent mechanical property prediction.

[0075] S3: using the prediction error as a reward function and adopting a deep deterministic policy gradient algorithm to construct a policy network.

[0076] Specifically, when constructing the policy network in S3, the method comprises:

[0077] S31: adopting a deep deterministic policy gradient algorithm to construct a policy network that dynamically adjusts the attention weight distribution of the feature prediction model according to the difference between the predicted performance and the target performance.

[0078] S32: defining the reward function of the policy network as the negative value of the prediction error:

[0079] R = -(|E pred -E true |+|σ pred -σ true |), wherein R is the reward function, E pred is the predicted elastic modulus, E true is the true elastic modulus, σ pred is the predicted compressive strength, and σ true is the true compressive strength; wherein the smaller the prediction error, the greater the reward value, guiding the policy network to output a better attention distribution scheme.

[0080] Wherein, when constructing the policy network, the state space, action space and reward function are designed; the state space design process is: at each time step t, the policy network receives the prediction performance error information of the current lattice structure as the state input:

[0081] s = (|E pred -E true |+|σ pred -σ true |); the action space design process is: the output of the policy network is taken as an action, which dynamically adjusts the attention weight (which can be additive or multiplicative adjustment) of each feature channel or position in the feature prediction model through the action, or the action can be directly applied to the original attention weight distribution.

[0082] By adopting the technical scheme, the policy network guides the model to converge to a direction with minimum performance error while dynamically adjusting the feature attention region, and a self-learning feedback closed-loop mechanism in the joint optimization process of structure generation and performance prediction is realized.

[0083] S4: jointly training the generative adversarial network model, the feature prediction model and the policy network to construct the lattice structure mechanical property prediction feedback model.

[0084] Specifically, when training the generative adversarial network model, the feature prediction model and the policy network in S4, the method comprises:

[0085] Generation phase: generating lattice structure parameters using the generative adversarial network model and predicting the mechanical properties of the candidate lattice structure using the feature prediction model to generate a prediction result.

[0086] Screening phase: screening samples with a preset mechanical property greater than a preset threshold, for example, elastic modulus > set threshold.

[0087] Feedback phase: updating the policy network parameters using reinforcement learning to optimize the generation and attention allocation strategy in the next round.

[0088] Pre-training phase: training the feature prediction model separately using a preset training data set and fixing the feature prediction model parameters.

[0089] Adversarial training phase: jointly optimizing the generator and discriminator of the generative adversarial network model and introducing a gradient penalty term to alleviate the training instability problem;

[0090] Reinforcement learning phase: fine-tuning the policy network parameters based on freezing the generator and the feature prediction model;

[0091] The training process is repeated until a preset number of iterations or a preset index is reached, and then the completed generative adversarial network model, feature prediction model and policy network are jointly trained to construct the lattice structure mechanical property prediction feedback model.

[0092] The method comprises: constructing a three-dimensional model of a random lattice structure according to the three-dimensional coordinates of a preset number of randomly generated unit nodes; calculating the stress-strain curve of the random lattice structure using finite element simulation; extracting the elastic modulus and compressive strength according to the stress-strain curve to form a training data set.

[0093] The optimization generation adversarial network model method comprises the following steps: initializing parameters of a generator and a discriminator; iteratively performing the following training steps: sampling from a real point lattice structure sample, sampling a virtual point lattice structure sample from a random noise vector, judging the real point lattice structure sample and the virtual point lattice structure sample by the discriminator, and updating the parameters of the discriminator to minimize the discriminator loss function; and fixing the discriminator and updating the parameters of the generator.

[0094] The method for training the policy network by using the DDPG reinforcement learning algorithm comprises the following steps: initializing the policy network and the evaluation network (including a main network and a target network), and initializing by using small random weights; iteratively performing the following training process: obtaining a prediction result from a current feature prediction model, calculating a state vector and inputting the state vector into the policy network, outputting an action and applying the action to a specified attention module in the feature prediction model, dynamically adjusting internal weights of the specified attention module, performing forward propagation again to obtain prediction performance of the updated feature prediction model, calculating a reward function according to the prediction and a real value and storing the reward function, the current state and the action in a replay pool, sampling a small batch of data by using an experience replay mechanism, updating a value function (state and action) according to the Bellman equation by using the evaluation network, and updating weights of the policy network by maximizing an expectation of the value function and synchronously updating parameters of the target network.

[0095] S5: inputting the node three-dimensional coordinates of the target point lattice structure into the point lattice structure mechanical property prediction feedback model to predict and output the mechanical property of the target point lattice structure.

[0096] In some embodiments of the present application, when the dynamic allocation of different position feature weights is introduced into the self-attention mechanism, the method comprises:

[0097] S131: calculating a normalized distance d i according to the node three-dimensional coordinate transformation, and calculating a sensitivity attenuation weight of each node three-dimensional coordinate position: wherein w i is the sensitivity attenuation weight of the i th node, d i is the normalized distance of the i th node, and a is a sensitivity attenuation factor for adjusting the attenuation intensity.

[0098] wherein when the normalized distance d i → 0 (the node is close to the lattice center), w i → 1, at this time, the weight is the highest; when the normalized distance d i → 1 (the node is far away from the lattice center), w at this time, the weight is reduced, which reflects that the edge contribution in the mechanical property is weak; the greater the value of a, the faster the attenuation, and the more prominent the central node.

[0099] S132: constructing a symmetric position sensitivity matrix: w j is the sensitivity attenuation weight of the jth node.

[0100] S133: the symmetric position sensitivity matrix is applied to the self-attention mechanism to realize a position sensitivity weighted attention mechanism: where ⊙ is the Hadamard product, used to dynamically adjust each attention score according to the sensitivity attenuation weights of the two nodes.

[0101] By adopting the technical solution, a distance-sensitive attenuation function combined with three-dimensional space geometric information is introduced, which fuses the normalized distance information of the three-dimensional coordinates of the nodes. In the feedback optimization of the mechanical properties of the point lattice structure, higher attention weights are given to the nodes close to the lattice center, the attention ability to the local key nodes (close to the lattice center region) is improved, the interference of the low-influence nodes far from the center region on the strategy network decision is suppressed, and therefore more accurate attention weight distribution is realized.

[0102] In some embodiments of the present application, when generating the generator of the optimization generative adversarial network model, the method comprises:

[0103] S401: calculate the Euclidean distances between all nodes and calculate the variance thereof as a node distribution uniformity index:

[0104] where N is the total number of nodes, is the average distance between all nodes, D ij is the Euclidean distance between the ith and jth nodes, D ij =‖x i -x j ‖2, x i is the three-dimensional coordinate of the ith node, x j is the three-dimensional coordinate of the jth node.

[0105] wherein the node distribution uniformity index aims to evaluate the discrete degree of the distance distribution between nodes, and a low discrete degree means more uniformity; the specific process is as follows: for the generated point lattice structure, there are N nodes and the Euclidean distance between nodes is defined as: D ij =‖x i -x j ‖2, then the average distance between all node pairs is calculated:

[0106] Then define the node distribution uniformity index U as the variance of these distances; wherein when U tends to 0, it indicates that the node distribution is more uniform.

[0107] S402: Re-arrange the dot array structure under the preset geometric transformation function and perform point set matching with the original dot array structure to calculate the configuration symmetry index: wherein T k is the geometric transformation function, S k is the average minimum matching error of the dot array structure under the k-th transformation with the original dot array structure, and K is the number of geometric transformations.

[0108] wherein the configuration symmetry index aims to evaluate the invariance degree of the generated structure under geometric transformation (such as mirror image, rotation), and reflects the symmetry of the configuration; the specific process is: setting K geometric transformation functions in the domain, such as rotation, flip, mirror image, and symmetric axis mapping, then performing each transformation on the generated structure to obtain a new structure point set, to measure the consistency between the original structure and the transformed structure, the minimum point pair error under each transformation is defined as:

[0109] The overall configuration symmetry index is wherein S k is smaller, indicating that the transformed dot array structure is closer to the original structure, and S is larger, indicating that the overall structure is more symmetrical.

[0110] S403: Add a weighted structure prior constraint term to the loss function of the generator to realize guided optimization of the uniformity and symmetry of the dot array structure parameters: wherein, is the total loss function of the generator, L adv is the adversarial loss function, λ U and λ S are hyperparameters for adjusting the weight of the structure prior constraint term in the total loss; wherein λ U is preferably in the range of [0.1, 5.0] to control the regularity of the generated node distribution; λ S is preferably in the range of [0.1, 10.0] to control the degree of geometric symmetry preservation.

[0111] By adopting the above technical solution, the structure prior constraint is added to the generative adversarial network to suppress the appearance of non-physical structures and improve the symmetry and uniformity of the generated structure.

[0112] In some embodiments of the present application, when optimizing the discriminator of the generative adversarial network model, the method further comprises:

[0113] S404: Input the dot array structure parameters generated by the generator and the real dot array structure parameters into the discriminator.

[0114] The lattice structure parameters generated by the generator are input into the discriminator together with the real lattice structure parameters, and the input format can adopt a splicing or parallel processing manner.

[0115] S405: The discriminator generates a discrimination output.

[0116] The discriminator output value represents the discrimination result of the real sample and the generated sample; when the Sigmoid activation function is used, the output range is [0, 1], which represents the probability of being a real sample.

[0117] S406: The total loss function of the discriminator is calculated by using a weighted combination of the adversarial loss function, the mean square error loss function and the structural similarity loss function.

[0118] The total loss function of the discriminator is: Wherein λ1, λ2, λ3 are the weight coefficients of the loss terms, which can be selected by cross-validation, and the commonly used value range is, for example, λ1 = 1.0, which retains the basic adversarial function; λ2 ∈ [0.1, 0.5], which improves the convergence stability; λ3 ∈ [0.01, 1.0], which is flexibly set according to the structural complexity. The adversarial loss function of the discriminator is The mean square error loss function of the discriminator is used to measure the error between the discriminator output and the target label. The structural similarity loss function of the discriminator is used to characterize the fidelity of the generated lattice structure in spatial distribution and topological characteristics.

[0119] S407: The discriminator parameters are updated according to the total loss function of the discriminator by using the gradient descent method or the Adam optimization method: Wherein η is the learning rate, ranging from 10 -5 to 10 -3 , the greater the value, the greater the update amplitude; The gradient (derivative vector) of the total loss function of the discriminator to its parameter θ D represents the fastest direction in which the loss function increases in the current direction.

[0120] By using the above technical solutions, when optimizing the discriminator, the weighted combination of the adversarial loss, the mean square error loss and the structural similarity loss is introduced, which can strengthen the sensitivity of the discriminator to the microstructure changes, improve its discrimination ability for the geometric distribution and local connection relationship between the generated sample and the real sample, and also prevent the discriminator from converging too fast or the gradient from disappearing, thereby providing more accurate quality evaluation signals for subsequent feedback.

[0121] Therefore, as shown in FIG. Figure 6 , the method comprises the following steps. Figure 6A module connection diagram of a hybrid architecture mechanical property prediction system for additive manufacturing lattice structures is shown. In some embodiments of the present application, to solve the above problems, the present application also designs a hybrid architecture mechanical property prediction system 2 for additive manufacturing lattice structures using the method described above, characterized by comprising: a feature extraction module 201 for converting the node three-dimensional coordinates of a plurality of lattice structures constructed randomly into normalized distances from the lattice center, and then introducing a self-attention mechanism and combining a residual connection and a layer normalization mechanism to construct a feature prediction model; a GAN construction module 202 for constructing a generative adversarial network model comprising a generator and a discriminator, wherein the generator takes a random noise vector as input and generates latent lattice structure parameters, and the discriminator utilizes a channel attention mechanism and combines an adversarial loss and a mean square error loss to optimize the feature recognition ability; a dynamic feedback module 203 for using the prediction error as a reward function and constructing a policy network using a deep deterministic policy gradient algorithm; a joint modeling module 204 for jointly training the generative adversarial network model, the feature prediction model and the policy network to construct a lattice structure mechanical property prediction feedback model; and a performance prediction module 205 for inputting the node three-dimensional coordinates of a target lattice structure into the lattice structure mechanical property prediction feedback model to predict the mechanical properties of the target lattice structure.

[0122] The present application establishes a learning model for predicting elastic modulus and compressive strength. This is mainly because elastic modulus mainly measures the deformation ability of materials in the elastic stage, reflecting the rigidity of the material; while compressive strength measures the maximum bearing capacity of the material under compression, reflecting the compressive performance of the material. These two have significant differences in physical meaning and form, so the prediction model needs to be optimized accordingly during modeling. At the same time, establishing separate models allows each model to be independently adjusted and optimized for hyperparameters, maximizing the prediction performance of each model. The effect of hyperparameters on model performance is shown in Figure 7 Figure 7 The effect of hyperparameters on model performance is shown in the machine learning of some embodiments of the present application, wherein Figure 7 (a) is the elastic modulus prediction model hyperparameter adjustment, Figure 7 (b) is the compressive strength model hyperparameter adjustment, Figure 7 (c) is the elastic modulus prediction model learning curve, Figure 7 (d) is the compressive strength prediction model learning curve. The machine learning prediction effect is excellent.

[0123] When applying the additive manufacturing lattice structure mechanical property prediction method or system of the present application, the comparison results of the elastic modulus prediction value and the true value are shown in Figure 8 The R 2 ​Score = 0.91, MSE = 150779; the predicted values of compressive strength are compared with the true values, and the results are shown in the figure Figure 9 The R 2 Score = 0.92, MSE = 8.86. The red line in the figure represents the function, and the points on the function represent the case where the predicted values are exactly equal to the true values. It can be seen that the points representing the predicted values of the elastic modulus and the compressive strength are close to the red line, which indicates that the machine learning prediction is accurate, and the key performance can be derived according to the design parameters of the lattice.

[0124] The results of the analysis of variance show that there is no significant relationship between the three-dimensional coordinates of the nodes and the mechanical properties, but the machine learning can accurately predict the mechanical properties according to the coordinate values, which indicates that there is a potential nonlinear relationship. According to the three-dimensional scatter diagram of the relationship between the elastic modulus and the node position Figure 10 The three-dimensional scatter diagram of the relationship between the compressive strength and the node position is shown in the figure Figure 11 It can be seen that the closer the node position is to the center (0.5), the higher the elastic modulus and the compressive strength are. As the node position approaches the edge (0 and 1), the elastic modulus and the compressive strength gradually decrease. At the same time, it can be clearly seen that the position of the node in the vertical direction has no significant effect on the elastic modulus and the compressive strength. This indicates that there is a local nonlinear effect in the data, and the analysis of variance is based on the global linear assumption, so the analysis of variance fails to detect this local effect.

[0125] In order to further analyze the significance of the influence of the node position on the mechanical properties, the three-dimensional coordinates of the nodes are converted. According to the analysis of the above Figure 10 and Figure 11 , the elastic modulus and the compressive strength decrease when the node position is far from the center of the lattice, so the distance of the three coordinate values of the node from the center point is taken as a factor, and the analysis of variance is performed again.

[0126] The results of the analysis of variance are shown in the following table:

[0127] Table: Analysis of variance of the distance factor of the node to the center of the lattice and the elastic modulus

[0128]

[0129] Table: Analysis of variance of the distance factor of the node to the center of the lattice and the compressive strength

[0130]

[0131] The analysis of variance shows that the horizontal X' and Y' position factors and their interaction (X':Y') have a significant effect on the compressive strength and elastic modulus (p value is much less than 0.05), while the vertical Z' position factor and its interaction with other factors have no significant effect. This shows that the position of the node in the horizontal direction has a significant positive effect on the compressive strength and elastic modulus, but the position in the vertical direction has a smaller effect.

[0132] The research analysis shows that using the method or system of the application can accurately predict the elastic modulus and compressive strength of the corresponding lattice structure through the three-dimensional coordinates of the nodes of the lattice structure. Using the prediction algorithm can make the degree of explanation of the independent variable to the dependent variable exceed 90%, with high prediction accuracy.

[0133] Therefore, the attention mechanism designed in the application significantly improves the ability of the model to capture local nonlinear effects (such as performance improvement when the node is close to the center). It is complementary to the results of variance analysis, and experiments show that the model based on the attention mechanism has higher prediction accuracy of the elastic modulus and compressive strength on the test set than the traditional ensemble learning method. The virtual structure generated by the GAN of the application is verified to have an error of less than 5% in the mechanical property distribution compared with the real data, verifying the effectiveness of the model. Therefore, the application realizes high-precision prediction and feedback of the mechanical properties of the lattice structure through the deep integration of the attention mechanism, the generative adversarial network and the policy network.

[0134] In some embodiments, with reference to Figure 12 as shown, Figure 12 a connection diagram of a terminal device for implementing the embodiments of the application is shown. The terminal device 3 includes a memory 301 and a processor 302, and the memory 301 stores a computer program executable on the processor 302. The processor 302 implements the method in the above embodiments when executing the computer program. The number of the memory 301 and the processor 302 can be one or more.

[0135] The terminal device 3 further includes:

[0136] a communication interface 303 for communicating with external devices and transmitting data.

[0137] If the memory 301, the processor 302 and the communication interface 303 are independently implemented, the memory 301, the processor 302 and the communication interface 303 can be connected to each other through a bus and complete communication between them.

[0138] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 12 Only one thick line is used in the figure to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0139] Optionally, in a specific implementation, if the memory 301, the processor 302, and the communication interface 303 are integrated on a chip, the memory 301, the processor 302, and the communication interface 303 can complete communication with each other through an internal interface.

[0140] The embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the program is executed by the processor 302 to implement the method provided in the embodiment of the present application.

[0141] The embodiment of the present application further provides a chip, which comprises the processor 302, is used for calling and running instructions stored in the memory 301 from the memory 301, so that a communication device installed with the chip executes the method provided in the embodiment of the present application.

[0142] The embodiment of the present application further provides a chip, which comprises an input interface, an output interface, a processor 302, and a memory 301, the input interface, the output interface, the processor 302, and the memory 301 are connected through an internal connection path, and the processor 302 is used for executing code in the memory 301, when the code is executed, the processor 302 is used for executing the method provided in the embodiment of the present application.

[0143] It is to be understood that the above-described processor 302 can be a central processing unit (CPU), but can also be other general purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, or the like. The general purpose processor can be a microprocessor or any conventional processor, or the like. It is to be noted that the processor 302 can be a processor supporting an advanced RISC machine (ARM) architecture.

[0144] Further, the above-described memory 301 can include read-only memory and random access memory, and can also include non-volatile random access memory. The memory 301 can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM) or flash memory. The volatile memory can include random access memory (RAM) used as an external cache. By way of example but not limitation, many forms of RAM are available. For example, static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM) and direct memory bus random access memory (Direct Rambus RAM, DRRAM).

[0145] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting the mechanical properties of a hybrid architecture for additively manufactured lattice structures, characterized in that: The method comprises: The three-dimensional coordinates of the nodes of the randomly constructed multiple lattice structures are converted into normalized distances from the lattice center, and then the self-attention mechanism is introduced and combined with the residual connection and layer normalization mechanism to build a feature prediction model; Construct a generative adversarial network model consisting of a generator and a discriminator, wherein the generator takes a random noise vector as input and generates latent lattice structure parameters, and the discriminator uses a channel attention mechanism and combines adversarial loss with mean squared error loss to optimize feature recognition capabilities; Use prediction error as the reward function and adopt deep deterministic policy gradient algorithm to build policy network; Combine and train the generative adversarial network model, feature prediction model, and policy network to construct a prediction and feedback model for the mechanical properties of lattice structures; The three-dimensional coordinates of the nodes of the target lattice structure are input into the lattice structure mechanical property prediction feedback model to predict and output the mechanical properties of the target lattice structure.

2. The method according to claim 1, characterized in that When constructing a feature prediction model, the method includes: randomly generating a preset number of node three-dimensional coordinates of a lattice structure and converting the node three-dimensional coordinates into normalized distances from the lattice center to highlight position sensitivity in the horizontal direction; Introducing interaction terms and high-order polynomial terms to enhance nonlinear representation; Introducing a self-attention mechanism that dynamically assigns weights to features at different positions: Among them, Q is the query matrix generated by linear mapping of the input features through the weight matrix, K is the keyword matrix representing the identification information of the concerned position generated by the linear mapping of the input features, and V is the value matrix representing the original information of each position generated by the linear mapping of the input features. QK T is the similarity score, d k is the scaling factor; and then generates the attention-weighted feature vector; The attention-weighted feature vector is input into a fully connected network and combined with residual connection and layer normalization mechanism to construct a feature prediction model for predicting mechanical properties.

3. The method according to claim 2, characterized in that When introducing a self-attention mechanism that dynamically assigns weights to features at different positions, the method includes: According to the normalized distance of the node's three-dimensional coordinate transformation, the sensitivity attenuation weight of each node's three-dimensional coordinate position is calculated: Among them, w i is the sensitivity attenuation weight of the i-th node, d i is the normalized distance of the i-th node, α is the sensitivity attenuation factor that adjusts the attenuation strength; Construct a symmetric position sensitivity matrix: w j is the sensitivity attenuation weight of the j-th node; The symmetric position sensitivity matrix is ​​applied to the self-attention mechanism to implement the position-sensitive weighted attention mechanism: Where ⊙ is the Hadamard product, which is used to dynamically adjust each attention score according to the sensitivity attenuation weights of the two nodes.

4. The method according to claim 2 or 3, characterized in that When constructing a generative adversarial network model including a generator and a discriminator, the method includes: Construct a generator that takes a random noise vector as input and generates the parameters of the latent lattice structure. In addition, a spatial attention mechanism is introduced to prioritize the horizontal node distribution during the generation process. Construct a discriminator that uses a channel attention mechanism that dynamically adjusts the weights of different feature channels and combines adversarial loss and mean squared error loss to optimize feature recognition capabilities; A generative adversarial network model is constructed based on the generator and the discriminator.

5. The method according to claim 4, characterized in that When constructing a policy network, the method includes: Using a deep deterministic policy gradient algorithm, we build a policy network that dynamically adjusts the attention weight allocation of the feature prediction model based on the difference between the predicted performance and the target performance. The reward function of the policy network is defined as the negative of the prediction error: R=-(|E pred -E true |+|σ pred -σ true |), where R is the reward function, E pred To predict the elastic modulus, E true is the true elastic modulus, σ pred To predict the compressive strength, σ true is the true compressive strength.

6. The method according to claim 5, characterized in that When training a generative adversarial network model, a feature prediction model, and a policy network, the method includes: Generate lattice structure parameters using a generative adversarial network model and predict the mechanical properties of the candidate lattice structure using a feature prediction model to generate a prediction result; Screening samples whose preset mechanical properties are greater than a preset threshold according to the prediction results; Update the policy network parameters and optimize the next round of generation and attention allocation strategies; Use the preset training data set to train the feature prediction model separately and fix the feature prediction model parameters; Jointly optimize the generator and discriminator of the generative adversarial network model and introduce gradient penalty terms to alleviate training instability issues; Based on the frozen generator and feature prediction model, fine-tune the policy network parameters; The training process is cyclical until the preset number of iterations or preset indicators are reached, and then the generative adversarial network model, feature prediction model and strategy network that have been trained are jointly constructed to construct a lattice structure mechanical property prediction feedback model.

7. The method according to claim 6, characterized in that The method for generating the preset training data set includes: Constructing a three-dimensional model of a random lattice structure according to the three-dimensional coordinates of a preset number of randomly generated unit nodes; The stress-strain curve of the random lattice structure was calculated using finite element simulation; The elastic modulus and compressive strength are extracted according to the stress-strain curve to form a training data set.

8. The method according to claim 6, characterized in that When optimizing a generator of a generative adversarial network model, the method includes: Calculate the Euclidean distance between all nodes and find its variance as an indicator of node distribution uniformity: Where N is the total number of nodes, is the average distance between all nodes, D ij is the Euclidean distance between the i-th and j-th nodes, D ij =‖x i -x j ‖2,x i is the three-dimensional coordinate of the i-th node, x j is the three-dimensional coordinate of the j-th node; The lattice structure is rearranged under the preset geometric transformation function and point set matching is performed with the original lattice structure to calculate the configurational symmetry index: Among them, T k is the geometric transformation function, S k is the average minimum matching error between the lattice structure and the original lattice structure under the kth transformation, K is the number of geometric transformations; A weighted structural prior constraint is added to the generator's loss function to achieve guided optimization of the uniformity and symmetry of the lattice structure parameters: in, is the total loss function of the generator, L adv To counter the loss function, λ U and λ S A hyperparameter for adjusting the weight of the structural prior constraint in the total loss.

9. The method according to claim 6 or 8, characterized in that When optimizing the discriminator of a generative adversarial network model, the method includes: Inputting the lattice structure parameters generated by the generator and the real lattice structure parameters into the discriminator; The discriminator generates a discriminative output; The total loss function of the discriminator is calculated using a weighted combination of the adversarial loss function, the mean square error loss function, and the structural similarity loss function; The discriminator parameters are updated using a gradient descent method or an Adam optimization method according to the total loss function of the discriminator.

10. A hybrid architecture mechanical properties prediction system for additive manufacturing lattice structures using the method according to any one of claims 1 to 9, characterized in that: include: The feature extraction module is used to convert the three-dimensional coordinates of the nodes of the randomly constructed multiple lattice structures into normalized distances from the lattice center, and then introduce a self-attention mechanism and combine residual connections and layer normalization mechanisms to build a feature prediction model; A GAN building module is used to construct a generative adversarial network model consisting of a generator and a discriminator, wherein the generator takes a random noise vector as input and generates latent lattice structure parameters, and the discriminator uses a channel attention mechanism and combines adversarial loss with mean squared error loss to optimize feature recognition capabilities; A dynamic feedback module that uses prediction error as a reward function and adopts a deep deterministic policy gradient algorithm to build a policy network; A joint modeling module is used to jointly train the generative adversarial network model, feature prediction model, and policy network to construct a prediction and feedback model for the mechanical properties of lattice structures; The performance prediction module is used to input the three-dimensional coordinates of the nodes of the target lattice structure into the lattice structure mechanical performance prediction feedback model to predict and output the mechanical properties of the target lattice structure.

Images