Neural network-based meshless point cloud three-dimensional structure topological optimization method and device

Abstract

The invention relates to the technical field of three-dimensional structure topological optimization, in particular to a meshless point cloud three-dimensional structure topological optimization method and device based on a neural network, and the method comprises the steps: reading a target graphic file, and determining a plurality of sampling points according to the target graphic file; based on a preset construction algorithm, constructing voxel units of a preset size for the plurality of sampling points, screening out target sampling points in a preset effective design domain based on the voxel units of the plurality of sampling points, and performing normalization processing on coordinates of the target sampling points to obtain processed sampling points; and inputting the processed sampling points into a preset neural network, and outputting an optimized three-dimensional structure topology distribution result. Therefore, the problems that topological optimization depends on finite element analysis, the calculation efficiency is low and the CAD model cannot be directly processed in the related technology are solved, and the efficiency of complex structure optimization and the engineering applicability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional structure topology optimization, and in particular relates to a meshless point cloud three-dimensional structure topology optimization method and device based on a neural network. BACKGROUND

[0002] Traditional topology optimization methods face challenges in automotive lightweight design, mainly relying on traditional techniques such as finite element analysis (FEA) and meshing, typical representatives such as the SIMP (Solid Isotropic Material with Penalization) method and the Level Set Method; for example, chassis component optimization requires repeated mesh reconstruction, resulting in a single component optimization cycle of up to 72 hours.

[0003] In recent years, methods based on data-driven and implicit representation of density field using networks have appeared in related technologies.

[0004] However, the methods in the related art rely on regular mesh partitioning and cannot handle cases where meshing is not applicable; FEA needs to be performed for each iteration of optimization, resulting in high computational cost; additional sensitivity analysis and filtering steps are required to avoid checkerboard phenomena or assist in training; the data-driven methods in the related art are highly dependent on the quality of training data, making it difficult to achieve target structure design with volume constraints, and unable to generate Pareto optimal structure solutions through controllable parameters; the neural network methods in the related art cannot achieve topology optimization of irregular design domains, cannot directly read geometric data from CAD models and construct neural network inputs, and cannot flexibly set material exclusion zones to meet complex engineering requirements, which need to be addressed. SUMMARY

[0005] The present application provides a meshless point cloud three-dimensional structure topology optimization method and device based on a neural network to solve the problems of topology optimization in related technologies relying on finite element analysis, low computational efficiency and inability to directly process CAD models.

[0006] The first aspect of the present application provides a meshless point cloud three-dimensional structure topology optimization method based on a neural network, comprising the following steps: reading a target graphics file and determining a plurality of sampling points based on the target graphics file; constructing a voxel unit of a preset size for each of the plurality of sampling points based on a preset construction algorithm, and filtering out target sampling points in a preset effective design domain based on the voxel units of the plurality of sampling points, and performing normalization processing on the coordinates of the target sampling points to obtain processed sampling points; Input the processed sampling points into a preset neural network to output an optimized three-dimensional structure topology distribution result.

[0007] Optionally, the preset neural network includes a preset displacement field neural network and a preset density field neural network.

[0008] Optionally, the preset displacement field neural network includes a Fourier feature mapping layer, a linear transformation layer, and a high-dimensional embedding layer.

[0009] Optionally, the preset displacement field neural network includes: The Fourier feature mapping layer projects the coordinate display of the processed sampling points into a high-dimensional frequency domain space. A first loss function of the preset displacement field neural network is constructed based on the principle of minimum potential energy, and a network outputting a static displacement is based on the first loss function and a preset displacement constraint condition.

[0010] Optionally, the preset density field neural network includes: A second loss function is constructed based on a preset compliance and a preset structure constraint condition, and the preset density field neural network is constructed based on the second loss function and a preset activation function.

[0011] Optionally, the first loss function is: ; Wherein, is the first loss function, is the internal strain energy, is the external force work.

[0012] Optionally, the second loss function is: ; Wherein, is the second loss function, is the initial compliance, is the current compliance, is the target volume fraction, is the current volume fraction, is the reward coefficient, is the penalty coefficient, is the number of points in the penalty area, is the current density of the i-th point.

[0013] The second aspect embodiment of the application provides a meshless point cloud three-dimensional structure topology optimization device based on a neural network, including: A reading module is configured to read a target graphic file and determine a plurality of sampling points based on the target graphic file.​ The processing module is configured to construct a voxel unit of a preset size for each of the plurality of sampling points based on a preset construction algorithm, filter out a target sampling point in a preset effective design domain based on the voxel units of the plurality of sampling points, and perform normalization processing on coordinates of the target sampling point to obtain a processed sampling point. The determining module is configured to input the processed sampling point into a preset neural network to output an optimized three-dimensional structure topology distribution result.

[0014] Optionally, the preset neural network includes a preset displacement field neural network and a preset density field neural network.

[0015] Optionally, the preset displacement field neural network includes a Fourier feature mapping layer, a linear transformation layer, and a high-dimensional embedding layer.

[0016] Optionally, the preset displacement field neural network includes: The Fourier feature mapping layer projects the coordinates of the processed sampling point into a high-dimensional frequency domain space. A first loss function of the preset displacement field neural network is constructed based on a principle of minimum potential energy, and a network outputting a static displacement is obtained based on the first loss function and a preset displacement constraint condition.

[0017] Optionally, the preset density field neural network includes: A second loss function is constructed based on a preset compliance and a preset structure constraint condition, and the preset density field neural network is constructed based on the second loss function and a preset activation function.

[0018] Optionally, the first loss function is: ; wherein, is the first loss function, is an internal strain energy, is an external force work.

[0019] Optionally, the second loss function is: ; wherein, is the second loss function, is an initial compliance, is a current compliance, is a target volume fraction, is a current volume fraction, is a reward coefficient, is a penalty coefficient, is a number of points in a penalty region, is a number of points in a reward region. a current density of the point.

[0020] The third aspect of the present application provides an electronic device, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the neural network-based meshless point cloud three-dimensional structure topology optimization method as described in the above embodiments.

[0021] The fourth aspect of the present application provides a computer readable storage medium having stored thereon a computer program, which is executed by a processor to implement the neural network-based meshless point cloud three-dimensional structure topology optimization method as described in the above embodiments.

[0022] Therefore, the embodiments of the present application read a target graphic file, and determine a plurality of sampling points according to the target graphic file; based on a preset construction algorithm, a voxel unit of a preset size is constructed for each of the plurality of sampling points, and a target sampling point in a preset effective design domain is screened out based on the voxel units of the plurality of sampling points, and the coordinates of the target sampling point are normalized to obtain a processed sampling point. The processed sampling point is input into a preset neural network, and an optimized three-dimensional structure topology distribution result is output. In this way, the problem that topology optimization in the related art relies on finite element analysis, has low calculation efficiency, and cannot directly process a CAD model is solved, and the efficiency and engineering applicability of complex structure optimization are improved.

[0023] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0024] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the accompanying drawings, wherein: Figure 1 A flowchart of a neural network-based meshless point cloud three-dimensional structure topology optimization method according to an embodiment of the present application is provided; Figure 2 A flowchart of a neural network-based meshless point cloud three-dimensional structure topology optimization method according to an embodiment of the present application is provided; Figure 3 A flowchart of a preset displacement field neural network and a preset density field neural network of a neural network-based meshless point cloud three-dimensional structure topology optimization method according to an embodiment of the present application is provided; Figure 4An output result schematic diagram of a neural network-based meshless point cloud three-dimensional structure topology optimization method according to an embodiment of the present application is a rectangular three-dimensional cantilever beam; Figure 5 An output result schematic diagram of a neural network-based meshless point cloud three-dimensional structure topology optimization method according to an embodiment of the present application is a rectangular structure with irregular multi-angle expansion area; Figure 6 An output result schematic diagram of a neural network-based meshless point cloud three-dimensional structure topology optimization method according to an embodiment of the present application is a larger hole model of a rectangular structure with irregular circular holes; Figure 7 An output result schematic diagram of a neural network-based meshless point cloud three-dimensional structure topology optimization method according to an embodiment of the present application is a smaller hole model of a rectangular structure with irregular circular holes; Figure 8 An output result schematic diagram of a neural network-based meshless point cloud three-dimensional structure topology optimization method according to an embodiment of the present application is a standard rectangular body with a material forbidden area; Figure 9 A schematic diagram of a neural network-based meshless point cloud three-dimensional structure topology optimization device according to an embodiment of the present application is shown in the figure; Figure 10 A structural schematic diagram of an electronic device according to an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0025] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0026] A neural network-based meshless point cloud three-dimensional structure topology optimization method and device of an embodiment of the present application are described below with reference to the accompanying drawings. In view of the problems of topology optimization in the related art relying on finite element analysis, low computational efficiency, and being unable to directly process CAD models, the present application provides a neural network-based meshless point cloud three-dimensional structure topology optimization method. In the method, an embodiment of the present application reads a target graphics file and determines a plurality of sampling points based on the target graphics file. Based on a preset construction algorithm, a voxel unit of a preset size is constructed for each of the plurality of sampling points. Based on the voxel units of the plurality of sampling points, a target sampling point in a preset effective design domain is filtered out. The coordinates of the target sampling point are normalized to obtain a processed sampling point. The processed sampling point is input to a preset neural network, and an optimized three-dimensional structure topology distribution result is output. Thus, the problem of topology optimization in the related art relying on finite element analysis, low computational efficiency, and being unable to directly process CAD models is solved, and the efficiency and engineering applicability of complex structure optimization are improved.

[0027] Specifically, Figure 1 A flowchart of a neural network-based meshless point cloud three-dimensional structure topology optimization method provided by an embodiment of the present application is shown in FIG. 1.

[0028] As Figure 1 shown, the neural network-based meshless point cloud three-dimensional structure topology optimization method includes the following steps: In step S101, a target graphics file is read, and a plurality of sampling points are determined based on the target graphics file. Specifically, the STP file of the target is read by OpenCASCADE, and a plurality of sampling points are determined based on the STP file.

[0029] In step S102, based on a preset construction algorithm, a voxel unit of a preset size is constructed for each of the plurality of sampling points. Based on the voxel units of the plurality of sampling points, a target sampling point in a preset effective design domain is filtered out. The coordinates of the target sampling point are normalized to obtain a processed sampling point.

[0030] The preset size can be a threshold value preset by the user, a threshold value obtained through a finite number of experiments, or a threshold value obtained through a finite number of computer simulations, which is not specifically limited here.

[0031] Specifically, a preset BRepExtrema_DistShapeShape algorithm is used to determine whether each sampling point is in the preset effective design domain by constructing a micro volume (cube or sphere) for each sampling point and determining whether the micro volume is in contact with the original geometry; the domain center is taken as the coordinate origin, and the coordinates of the target sampling points that pass the screening are normalized so that the coordinate range of the longest dimension is from -0.5 to 0.5, which is in the sensitive numerical range of the neural network, and the training is more stable.

[0032] In step S103, the processed sampling points are input into the preset neural network, and an optimized three-dimensional structure topology distribution result is output.

[0033] Specifically, the sampling points that pass the effective domain screening and normalization processing are input into the preset neural network to generate a three-dimensional structure topology distribution result that meets the volume constraint and the individualized forbidden area requirement.

[0034] Optionally, in some embodiments, the preset neural network includes a preset displacement field neural network and a preset density field neural network.

[0035] It can be understood that the STP file is read by OpenCASCADE, and the minimum bounding box and vertex coordinates thereof are extracted, and random or regular sampling points are generated inside the bounding box. The regular sampling points are evenly distributed in the set. The number of sampling points is preferably set to about 5000, which is used for input of the density neural network and calculation of the volume fraction constraint and input of the displacement neural network to solve the static response.

[0036] Optionally, in some embodiments, the preset displacement field neural network includes a Fourier feature mapping layer, a linear transformation layer, and a high-dimensional embedding layer.

[0037] Optionally, in some embodiments, the preset displacement field neural network includes a Fourier feature mapping layer that projects the coordinates of the processed sampling points into a high-dimensional frequency domain space; a first loss function of the preset displacement field neural network is constructed based on the principle of minimum potential energy, and a network that outputs a static displacement based on the first loss function and a preset displacement constraint condition.

[0038] Optionally, in some embodiments, the first loss function is: ; wherein, is the first loss function, is the internal strain energy, is the external force work.

[0039] Understandably, the first layer of the pre-defined displacement field neural network uses a Fourier feature mapping layer inspired by "Fourier features let networks learn high frequency functions in low dimensional domains" to project the coordinate display onto a high-dimensional frequency domain space. The second layer is a linear transformation layer used to output the three-dimensional displacement field. The displacement field prediction result is as follows: ; in, For the displacement field prediction results, This represents the input set of coordinate points. The weight matrix has a shape of (3, kernel size). As a paranoia vector, This is also a weight matrix with a shape of (kernel size, 3). The neural network structure preferably employs high-dimensional embedding layers to enhance coordinate representation capabilities; displacement constraints are added to the output, with values ​​set to 0 at fixed interfaces. The preset loss function of the displacement field neural network uses the minimum potential energy principle, enabling the neural network to learn the weights of the displacement that minimizes the potential energy of its output, thus obtaining the network with the output static displacement; the displacement neural network loss function (i.e., the first loss function) The design is derived from Young's modulus (1000.0), Poisson's ratio (0.3), global and boundary condition regions, stress matrix, displacement information of points within the domain, and automatic differentiation calculations in Tensorflow. This design significantly improves generalization by driving training through physical equations rather than supervised data.

[0040] Optionally, in some embodiments, the preset density field neural network includes: constructing a second loss function based on a preset flexibility and preset structural constraints, and constructing a preset density field neural network based on the second loss function and a preset activation function.

[0041] Optionally, in some embodiments, the second loss function is: ; in, For the second loss function, For initial flexibility, For the current flexibility, For the target volume fraction, This represents the current volume fraction. As the reward coefficient, The penalty coefficient is... The number of points in the penalty area. For the first The current density of each point.

[0042] It can be understood that the loss function of the density neural network (second loss function) On the basis of the traditional volume constraint, the flexibility is ensured to be as small as possible, and a personalized structure constraint, i.e., a density penalty of a prohibited material region, is introduced. The network structure of the preset density field neural network is similar to that of the preset displacement field neural network, and the output does not need to add displacement constraints, but adds a sigmoid activation function (preset activation function) to ensure that the output value is between 0-1, which identifies the density corresponding to the coordinate point of the output. The density field output result is , wherein, is the preset activation function.

[0043] It should be noted that when the preset displacement field neural network and the preset density field neural network are integrated and trained, first, the coordinates sampled randomly are used for forward propagation through the density network and the displacement network, and the displacement network is trained with the current density output and the boundary condition. In the initial stage, the displacement neural network can be pre-trained for several rounds (such as 1000 times) to converge to an approximate static equilibrium state, and thereafter, since the overall structure changes little, only a small step update (such as 20 times) is needed in each round of topological iteration to maintain the response accuracy. The output of the preset displacement field neural network and the current density can be passed to the density network loss function, wherein the part for calculating the volume fraction constraint needs to be calculated using regular sampling points to avoid significant fluctuations in the volume fraction in each iteration. In the entire training process, the optimization rounds can be set in the range of 100-1000, the learning rate is set to 5.0*10-6, and both use the Adam optimizer.

[0044] Therefore, the present application solves the problem of high dependence on mesh, finite element analysis and sensitivity analysis in the process of topological optimization of part structures in the fields of automobiles, aerospace, etc., and cannot directly read from the CAD model file and realize personalized end-to-end topological optimization. By automatically constructing the point cloud domain data of the geometry, inputting into the displacement neural network and the density neural network, replacing FEA with physical information as the guide, constructing a personalized structure constraint control mechanism, and then realizing end-to-end personalized topological optimization of three-dimensional structures with irregular design domains. The present application greatly improves the automation, intelligence and engineering practicability of topological optimization, and is suitable for application scenarios such as complex structural parts, customized functional parts, etc., such as weight reduction topological optimization of body-in-white structural parts (A / B pillars, longitudinal beams, etc.), and generation of special-shaped structures of interior functional parts.

[0045] To make the person skilled in the art further understand the neural network-based meshless point cloud three-dimensional structure topological optimization method of the embodiments of the present application, the embodiments will be described in detail below with reference to the accompanying drawings. Figures 2 to 8

[0046] Specifically, as shown in Figure 2 Figure 2 ​​A flowchart of a neural network-based meshless point cloud three-dimensional structure topology optimization method for an embodiment of the present application; the neural network-based meshless point cloud three-dimensional structure topology optimization method is divided into the following parts: three-dimensional geometry reading and point cloud construction, effective field determination and coordinate transformation, displacement neural network construction using displacement constraint and minimum potential energy loss function, density neural network construction with personalized structure constraint, and displacement and density neural network integrated training. Among them, as shown in Figure 3 , Figure 3 A design flowchart of a preset displacement field neural network and a preset density field neural network of a neural network-based meshless point cloud three-dimensional structure topology optimization method for an embodiment of the present application; according to the shape of the geometric body, point coordinates are randomly or regularly generated in the design domain, and boundary condition point coordinates are determined, which are used as input data. On the one hand, displacement output and potential energy are calculated through the displacement neural network; on the other hand, density output is obtained through the density neural network, which is used for loss function calculation; the results of the two paths jointly act on the loss function, and after iterative optimization, the output result is the optimized three-dimensional structure topology, realizing efficient and accurate optimization design of complex geometric bodies under the condition of meshless point cloud.

[0047] Therefore, the present application reads set data from the STP file, constructs the optimization domain through point cloud generation, normalization and other methods, and realizes structure response prediction and material personalized distribution optimization through the density field neural network and the displacement field neural network, which is suitable for various engineering application scenarios. This method significantly improves the automation, universality and engineering practicability of the topology optimization system, and is suitable for the optimization design of complex irregular structures in automobile part design.

[0048] Further, as shown in Figure 4 , Figure 4The output result schematic diagram of a neural network-based meshless point cloud three-dimensional structure topology optimization method of one embodiment of the present application is a regular rectangular three-dimensional cantilever beam. The present embodiment selects a regular rectangular three-dimensional cantilever beam as the research object. The geometry is generated by CATIA (Computer-Aided Three-dimensional Interactive Application) design, with length, width and height set to 40, 8 and 20 respectively, and exported as an STP format. The OpenCASCADE module is used to read the STP file, and point cloud data is generated by regular and random sampling methods, with 6,000 random sampling points and 4,536 regular sampling points. The constraint loading is set to fixed support on the left side and downward concentrated force on the right side. The above point cloud data and boundary conditions are input into the trained topology optimization neural network model to generate the optimized structure distribution result. All experiments are performed on a device. The device is equipped with a 13th generation Intel (R) Core (TM) i7-13700H 2.40GHz processor, 32G memory and an Intel GeForce RTX 4070 graphics card. Due to device limitations and training time limitations, the present embodiment omits the random sampling point generation and screening process each time and uses the generated regular sampling points and screening instead. The density neural network optimization iteration times are set to 100, and still better results are obtained. If a better device is used or the training rounds are increased, the final result will still have a lot of room for improvement.

[0049] Further, as shown in Figure 5 , Figure 5 The output result schematic diagram of a neural network-based meshless point cloud three-dimensional structure topology optimization method of one embodiment of the present application is a rectangular structure with an irregular multi-angle expansion area. The present embodiment selects a rectangular structure with an irregular multi-angle expansion area as the research object. A trapezoidal protruding area is added to the upper right of the original regular rectangular body, increasing the overall height by 10 and the highest point to 30. The original rectangular body size remains 40 in length, 8 in width and 20 in base height. After modeling in CATIA, the whole is exported as an STP format. The OpenCASCADE module is used to read the STP model and generate point clouds, and the minimum bounding box size obtained is 40x8x30. The random sampling generates 4,752 points, and the regular sampling generates 4,608 points. The left side is set as a fixed boundary, and the right side is subjected to a vertical downward concentrated force. After building the constraint input, it is transmitted into the training model for topology optimization to obtain the structure distribution result.

[0050] Further, as shown in Figure 6 and Figure 7 , Figure 6This illustration shows the output results of a large-hole model of a rectangular structure with irregular circular holes, based on a neural network-based meshless point cloud 3D structure topology optimization method according to one embodiment of this application. Figure 7 This embodiment of the present application presents a method for topology optimization of a meshless point cloud 3D structure based on a neural network. The research object is a schematic diagram of the output result of a smaller hole model of a rectangular structure with irregular circular holes. This embodiment studies the topology optimization problem of a rectangular structure with irregular circular holes. The geometry is designed in CATIA as a basic rectangular body with dimensions of 40 (length), 8 (width), and 20 (height), and one or more irregular circular hole regions are added inside the structure. After exporting to an STP file, invalid points within the holes can be automatically removed during the point cloud filtering process. The larger hole model has 3976 regular sampling points and 5222 random sampling points; the smaller hole models have 4410 and 5820 points respectively. The left end of the structure is set as a fixed boundary, and a vertical concentrated force is applied to the right end. All information is input into the network model to complete training and optimization.

[0051] Furthermore, such as Figure 8 As shown, Figure 8 This embodiment of the present application presents a method for topology optimization of a meshless point cloud 3D structure based on a neural network, using a standard rectangular prism with a material forbidden zone as the research object. This embodiment verifies the invention's ability to personalize the "material forbidden zone." The structural input is a standard rectangular prism with dimensions of 40×8×20. During CATIA modeling, the upper right corner of the geometry is preset as a "forbidden zone" where materials are not desired. After point cloud sampling and coordinate filtering, the points in this zone are automatically marked and additional penalty weights are applied to the loss function of the density neural network. The sampling point settings are consistent with Embodiment 1: 6000 random points and 4536 regular points. Loading constraints are set and... Figure 4 The same implementation method is used, and after inputting the above information into the network, an optimized structural result that avoids the distribution of materials in the restricted area is obtained.

[0052] Therefore, the application proposes a neural network-based end-to-end structure topology optimization method, including but not limited to STP geometry reading, point cloud generation, neural network construction and integrated training, etc. Through physical equation instead of supervised data-driven training, the method replaces the finite element calculation mode, and significantly improves the adaptability and engineering adaptability of the topology optimization process. The application proposes to read CAD files, automatically generate point clouds, and generate data that can be directly input into the topology optimization neural network through a geometry distance judgment method, solving the difficulty of lacking direct connection between CAD models and neural network topology optimization. The application realizes direct mapping between CAD models and neural network inputs through point cloud generation and geometry distance determination algorithm, combined with re-marking boundary conditions and coordinate transformation steps, realizes reading and optimization calculation for irregular design domains, reduces the overhead and dependence of mesh division, and solves the problem that topology optimization cannot directly access original engineering data in the prior art. The application constructs a personalized structure constraint control mechanism, adds a density penalty term of material disabled area in the loss function, realizes controllability and personalized design capability of structure optimization.

[0053] According to the neural network-based meshless point cloud three-dimensional structure topology optimization method proposed in the embodiments of the application, a target graphic file is read, and a plurality of sampling points are determined according to the target graphic file; based on a preset construction algorithm, a voxel unit of a preset size is constructed for each of the plurality of sampling points, and a target sampling point in a preset effective design domain is filtered out based on the voxel units of the plurality of sampling points, and the coordinates of the target sampling point are normalized to obtain a processed sampling point. The processed sampling point is input into a preset neural network, and an optimized three-dimensional structure topology distribution result is output. Thus, the problem of topology optimization relying on finite element analysis, low calculation efficiency and being unable to directly process CAD models in related technologies is solved, and the efficiency and engineering applicability of complex structure optimization are improved.

[0054] Secondly, the neural network-based meshless point cloud three-dimensional structure topology optimization device proposed in the embodiments of the application is described with reference to the accompanying drawings.

[0055] Figure 9 is a block schematic diagram of the neural network-based meshless point cloud three-dimensional structure topology optimization device in the embodiments of the application.

[0056] As shown in Figure 9 , the neural network-based meshless point cloud three-dimensional structure topology optimization device 1000 includes a reading module 100, a processing module 200 and a determination module 300.

[0057] The reading module 100 is configured to read a target graphic file and determine a plurality of sampling points according to the target graphic file; The processing module 200 is configured to construct a preset size of a voxel unit for each of the plurality of sampling points based on a preset construction algorithm, filter out a target sampling point in a preset effective design domain based on the voxel units of the plurality of sampling points, and perform normalization processing on coordinates of the target sampling point to obtain a processed sampling point. The determining module 300 is configured to input the processed sampling point into a preset neural network to output an optimized three-dimensional structure topology distribution result.

[0058] Optionally, the preset neural network includes a preset displacement field neural network and a preset density field neural network.

[0059] Optionally, the preset displacement field neural network includes a Fourier feature mapping layer, a linear transformation layer, and a high-dimensional embedding layer.

[0060] Optionally, the preset displacement field neural network includes that the Fourier feature mapping layer projects the coordinates of the processed sampling point into a high-dimensional frequency domain space; a first loss function of the preset displacement field neural network is constructed based on a principle of minimum potential energy, and a network outputting a static displacement is based on the first loss function and a preset displacement constraint condition.

[0061] Optionally, the preset density field neural network includes that a second loss function is constructed based on a preset compliance and a preset structure constraint condition, and the preset density field neural network is constructed based on the second loss function and a preset activation function.

[0062] Optionally, the first loss function is as follows: ; wherein, is the first loss function, is an internal strain energy, is an external force work.

[0063] Optionally, the second loss function is as follows: ; wherein, is the second loss function, is an initial compliance, is a current compliance, is a target volume fraction, is a current volume fraction, is a reward coefficient, is a penalty coefficient, is a number of points in a penalty region, is a current density of the i-th point in the penalty region.

[0064] ​It should be noted that the foregoing explanation and description of the embodiment of the neural network-based meshless point cloud three-dimensional structure topology optimization method also applies to the neural network-based meshless point cloud three-dimensional structure topology optimization device of this embodiment, which will not be described here.

[0065] According to the neural network-based meshless point cloud three-dimensional structure topology optimization device provided in the embodiment of the present application, the target graphic file is read, and a plurality of sampling points are determined according to the target graphic file; based on a preset construction algorithm, a voxel unit of a preset size is constructed for each of the plurality of sampling points, and a target sampling point in a preset effective design domain is screened out based on the voxel units of the plurality of sampling points, and the coordinates of the target sampling point are normalized to obtain a processed sampling point. The processed sampling point is input into a preset neural network, and an optimized three-dimensional structure topology distribution result is output. Thus, the problem of topology optimization relying on finite element analysis, low calculation efficiency and inability to directly process CAD models in related technologies is solved, and the efficiency and engineering applicability of complex structure optimization are improved.

[0066] Figure 10 The electronic device provided in the embodiment of the present application is shown in the structural schematic diagram. The electronic device can include: The memory 1001, the processor 1002, and the computer program stored in the memory 1001 and executable on the processor 1002.

[0067] The processor 1002 implements the neural network-based meshless point cloud three-dimensional structure topology optimization method provided in the above embodiments when executing the program.

[0068] Further, the electronic device further includes: The communication interface 1003 is used for communication between the memory 1001 and the processor 1002.

[0069] The memory 1001 is used to store the computer program executable on the processor 1002.

[0070] The memory 1001 can include a high-speed RAM memory, and can also include a non-volatile memory, such as at least one disk memory.

[0071] If the memory 1001, the processor 1002 and the communication interface 1003 are implemented independently, the communication interface 1003, the memory 1001 and the processor 1002 can be connected with each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (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 convenience of representation, Figure 10 Only one thick line is used to represent the bus in the figure, but it does not mean that there is only one bus or only one type of bus.

[0072] Optionally, in a specific implementation, if the memory 1001, the processor 1002 and the communication interface 1003 are integrated on a chip, the memory 1001, the processor 1002 and the communication interface 1003 can complete communication between each other through an internal interface.

[0073] The processor 1002 can be a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement one or more embodiments of the present application.

[0074] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the neural network based meshless point cloud three-dimensional structure topology optimization method.

[0075] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0076] Furthermore, the terms "first", "second", etc. are used herein only to describe different steps or features and do not imply a relative importance or a specific order of steps or features. Thus, features defined with "first", "second" etc. can include at least one of the features implicitly or explicitly. In the description of the application, the meaning of "N" is at least two, such as two, three, etc., unless otherwise expressly specified.

[0077] Any process or method descriptions or blocks in flow charts described herein and elsewhere can be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process. Alternate implementations are included within the scope of the preferred embodiments of the present application in which the functions performed by the various processes described herein and elsewhere are allocated differently among the components of the preferred embodiments, such as according to the functions performed by the various components, in a substantially simultaneous manner, or according to a different order.

[0078] It should be understood that aspects of the present application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, the various steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. As such, if implemented in hardware and in another embodiment, the hardware can include any or a combination of the following technologies, which are all well known in the art: a discrete logic circuit having logic gates for implementing logic functions upon an application of data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0079] Those of skill in the art would understand that information and signals can be represented using any of a variety of technologies and techniques. Further, those of skill in the art would understand that the preferred embodiments of the present application can be implemented by a variety of means, including as a process, an apparatus, a machine, or a combination thereof. In addition, the preferred embodiments of the present application can be implemented as a software program or other computer instructions implemented on a machine such as a computer or any other programmable control device. As such, the design and structure of each hardware component, as well as the interactions between the hardware components, can be varied, depending on the particular implementation.

Claims

1. A method for topology optimization of meshless point cloud 3D structures based on neural networks, characterized in that, Includes the following steps: Read the target image file and determine multiple sampling points based on the target image file; Based on a preset construction algorithm, voxel units of a preset size are constructed for multiple sampling points respectively, and target sampling points in a preset effective design domain are selected based on the voxel units of multiple sampling points. The coordinates of the target sampling points are then normalized to obtain the processed sampling points. The processed sampling points are input into a preset neural network, which outputs the optimized three-dimensional structural topology distribution result.

2. The method according to claim 1, characterized in that, The preset neural network includes a preset displacement field neural network and a preset density field neural network.

3. The method according to claim 2, characterized in that, The preset displacement field neural network includes: a Fourier feature mapping layer, a linear transformation layer, and a high-dimensional embedding layer.

4. The method according to claim 3, characterized in that, The preset displacement field neural network includes: The Fourier feature mapping layer displays the coordinates of the processed sampling points by high-dimensional frequency domain spatial projection. The first loss function of the preset displacement field neural network is constructed based on the principle of minimum potential energy, and the network outputs static displacement based on the first loss function and the preset displacement constraint conditions.

5. The method according to claim 2, characterized in that, The preset density field neural network includes: A second loss function is constructed based on a preset flexibility and a preset structural constraint, and the preset density field neural network is constructed based on the second loss function and a preset activation function.

6. The method according to claim 4, characterized in that, The first loss function is: ； in, Let the first loss function be... As internal strain energy, This is work done by external force.

7. The method according to claim 5, characterized in that, The second loss function is: ； in, For the second loss function, For initial flexibility, For the current flexibility, For the target volume fraction, This represents the current volume fraction. As the reward coefficient, The penalty coefficient is... The number of points in the penalty area. For the first The current density of each point.

8. A topology optimization device for a meshless point cloud 3D structure based on a neural network, characterized in that, include: The reading module is used to read the target graphic file and determine multiple sampling points based on the target graphic file; The processing module is used to construct voxel units of a preset size for multiple sampling points based on a preset construction algorithm, and to filter out target sampling points in a preset valid design domain based on the voxel units of multiple sampling points, and to normalize the coordinates of the target sampling points to obtain the processed sampling points. The determination module is used to input the processed sampling points into a preset neural network and output the optimized three-dimensional structure topology distribution result.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the neural network-based meshless point cloud 3D structure topology optimization method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the neural network-based meshless point cloud 3D structure topology optimization method as described in any one of claims 1-7.

Images