Deep learning driven 3d parametric design method and system for hydraulic engineering

Abstract

The application relates to the field of artificial neural network architecture and three-dimensional parameterization design technology, and discloses a deep learning driven three-dimensional parameterization design method and system for hydraulic engineering. The method constructs and runs a deep learning network topology, wherein the deep learning network topology comprises a shared topology address coding layer, an address hard routing double-flow graph convolution layer, a closed loop tensor constraint layer, a boundary conflict gate projection layer, a loop modulation cross attention layer and a boundary backfill three-dimensional decoding layer which are connected in sequence; three-dimensional parameterization design data of the hydraulic engineering is coded into a node boundary tensor, an edge adjacency tensor, a parameter control tensor and a boundary mask tensor, and a three-dimensional parameterization geometry matrix and a topology assembly diagram are generated through the above network layers. The application is used for structurally transmitting geometry topology, hydraulic boundary, parameter control and loop constraint in the deep learning network architecture.

Description

Technical Field

[0001] This invention relates to the field of artificial neural network architecture and three-dimensional parametric design data processing technology, and in particular to a deep learning-driven three-dimensional parametric design method and system for water conservancy projects. Background Technology

[0002] With the application of deep learning in 3D geometric modeling, parametric design, and engineering structure simulation, 3D parametric design data is gradually being transformed from traditional explicit geometric modeling data into tensor, graph structure, or mesh-based feature representations that can be processed by neural networks. In the 3D parametric design scenario of hydraulic engineering, the design object typically contains multiple types of information, including geometric boundaries, structural contours, parametric control points, spatial connectivity relationships, and hydraulic boundaries. This information possesses both continuous geometric features and discrete topological connectivity features in its data structure, and therefore can be abstracted into a 3D topology generation and parametric representation problem with multi-source constraints.

[0003] In existing technologies, deep learning processing methods for 3D CAD objects typically first convert the boundary representation data of the 3D model into a graph structure or mesh feature suitable for machine learning models. For example, patent application CN115146333A discloses a technical solution for generating UV-Net representations of 3D CAD objects. This solution generates a graph structure from the boundary representation of the 3D CAD object and extracts geometric features based on the UV mesh of parametric surfaces, enabling machine learning models to simultaneously process the geometric and topological information of the 3D object. This type of technical solution can provide a relatively standardized data representation foundation for tasks such as 3D CAD object recognition, classification, and similarity analysis.

[0004] However, in the three-dimensional parametric design data of hydraulic engineering projects, there are usually stronger constraint relationships between geometric parameters, hydraulic boundaries, structural partitions, and assembly connections. Existing deep learning network architectures based on graph structures or UV meshes mainly focus on transforming geometric patches, boundary curves, and adjacency relationships into a unified input representation. These networks typically employ general graph convolution, convolutional neural networks, or attention networks for feature propagation. In this approach, geometric features and hydraulic boundary features are easily mixed and propagated within the same feature channel, leading to coupling interference between constraint information from different sources during node updates. Simultaneously, ordinary graph convolutional networks primarily rely on adjacency edges for local message passing, lacking explicit constraints on the closed topological relationships formed by boundary loops, cross-sectional loops, and assembly loops. This can easily lead to inconsistencies in surface patch connections, local topological breaks, or mismatched loop relationships during the three-dimensional parametric geometry generation process.

[0005] Furthermore, existing deep learning 3D generative networks typically use boundary markers as input features or post-processing conditions after decoding when handling multiple types of boundary conditions. When hydraulic boundaries, geometric boundaries, and structural connection boundaries intersect at the same spatial node or adjacent topological nodes, the network lacks a structured processing mechanism to distinguish and propagate boundary constraint conflicts, leading to inconsistent propagation of boundary features during the fusion and decoding stages. Particularly when 3D parametric geometric matrices and topological assembly diagrams need to be generated simultaneously, if closed-loop constraints, boundary constraints, and parameter control constraints cannot maintain a consistent data flow relationship during the forward propagation of the deep learning network, it will affect the correspondence between the 3D geometric representation and the topological assembly relationship.

[0006] Therefore, the main technical problem that urgently needs to be solved in the three-dimensional parametric design of hydraulic engineering based on deep learning is: how to maintain a consistent structured transmission relationship of various constraint information such as geometric topology, hydraulic boundary, parameter control and assembly loop during the forward propagation process within the deep learning network architecture, so as to reduce the impact of multi-source feature coupling, mismatch of closed topology relationship and boundary constraint conflict on the generation of three-dimensional parametric geometry from the bottom layer of the algorithm architecture. Summary of the Invention

[0007] To overcome the aforementioned technical deficiencies, the present invention aims to provide a deep learning-driven three-dimensional parametric design method and system for hydraulic engineering. The present invention employs a deep learning network topology sequentially connected by a shared topology address encoding layer, an address hard-routing dual-flow graph convolutional layer, a closed-loop tensor constraint layer, a boundary conflict gated projection layer, a loop modulation cross-attention layer, and a boundary backfeed three-dimensional decoding layer. This allows multiple types of constraint information, such as geometric topology, hydraulic boundaries, parameter control, and assembly loops, to be encoded, routed, constrained, fused, and decoded under the same shared topology address. This addresses the problems of multi-source feature coupling, closed-loop topology mismatch, and boundary constraint conflicts from the bottom layer of the algorithm architecture.

[0008] This invention discloses a deep learning-driven three-dimensional parametric design method for hydraulic engineering. It constructs and runs a deep learning network topology, which includes a shared topology address encoding layer, an address hard-routing dual-flow graph convolutional layer, a closed-loop tensor constraint layer, a boundary conflict-gated projection layer, a loop modulation cross-attention layer, and a boundary backfeed three-dimensional decoding layer, all connected sequentially. The method includes: Shared topology addresses are assigned to the three-dimensional parametric design data of water conservancy projects, and the three-dimensional parametric design data of water conservancy projects are encoded into node boundary tensors, edge adjacency tensors, parameter control tensors and boundary mask tensors carrying the same shared topology address; The shared topology address coding layer generates node address embeddings, node-edge routing tables, node-parameter routing tables, and node-loop routing tables. A shared edge-gated matrix, geometric flow node features, and hydraulic flow node features are generated by a dual-flow graph convolutional layer based on address hard routing. Loop constraint tensors are generated from closed-loop tensor constraint layers; Boundary gated features are generated from the boundary conflict-gated projection layer; The attention weights are modulated by the loop-modulated cross-attention layer using loop constraint tensors to generate a fused latent vector; The 3D parametric geometry matrix and topology assembly diagram are generated by the boundary backfilling 3D decoding layer.

[0009] Preferably, the shared topology address includes node address, edge address, loop address, and parameter control address; the node boundary tensor carries the node address, the edge adjacency tensor carries the edge address, the parameter control tensor carries the parameter control address, and the boundary mask tensor carries the node address and loop address.

[0010] Preferably, the shared topology address coding layer includes an address allocation sublayer, an address alignment sublayer, and an address routing table generation sublayer; the address allocation sublayer allocates shared topology addresses to nodes, edges, loops, and parameter control points respectively; the address alignment sublayer aligns the node boundary tensors, boundary mask tensors, and parameter control tensors with the same node addresses; and the address routing table generation sublayer generates node-edge routing tables, node-parameter routing tables, and node-loop routing tables.

[0011] Preferably, the shared topology address coding layer further includes a suspended address isolation sublayer; the suspended address isolation sublayer writes the shared topology address that has not formed a routing connection in the node-edge routing table, node-parameter routing table or node-loop routing table into the suspended address channel, where the shared topology address is a node address, edge address, loop address or parameter control address, and the output of the suspended address channel is connected to the address hard routing dual-flow graph convolutional layer.

[0012] Preferably, the address hard routing dual-flow graph convolutional layer includes a geometric flow graph convolutional branch, a hydraulic flow graph convolutional branch, and a shared edge gating sublayer; the shared edge gating sublayer generates a shared edge gating matrix based on the node-edge routing table, the node-parameter routing table, and the boundary mask tensor, and both the geometric flow graph convolutional branch and the hydraulic flow graph convolutional branch receive the shared edge gating matrix.

[0013] Preferably, the address hard routing dual-flow graph convolutional layer further includes a cross-flow residual exchange sublayer; the cross-flow residual exchange sublayer includes a geometry-to-hydraulic exchange channel and a hydraulic-to-geometric exchange channel, the geometry-to-hydraulic exchange channel writes the geometric flow node features into the hydraulic flow graph convolutional branch according to the shared edge gating matrix, and the hydraulic-to-geometric exchange channel writes the hydraulic flow node features into the geometric flow graph convolutional branch according to the shared edge gating matrix.

[0014] Preferably, the closed loop tensor constraint layer includes a boundary loop extraction sublayer, a cross-sectional loop extraction sublayer, an assembly loop extraction sublayer, and a loop tensor generation sublayer; the boundary loop extraction sublayer extracts boundary loop tensors based on the node-loop routing table, the cross-sectional loop extraction sublayer extracts cross-sectional loop tensors based on edge adjacency tensors, the assembly loop extraction sublayer extracts assembly loop tensors based on parameter control tensors, and the loop tensor generation sublayer generates loop constraint tensors based on the boundary loop tensors, cross-sectional loop tensors, and assembly loop tensors.

[0015] Preferably, the closed loop tensor constraint layer further includes a loop direction encoding sublayer; the loop direction encoding sublayer generates a loop direction tensor according to the arrangement order of the shared topology addresses corresponding to adjacent edges in the same loop, and writes the loop direction tensor into the loop constraint tensor.

[0016] Preferably, the boundary conflict gating projection layer includes a hydraulic boundary gating sublayer, a geometric boundary gating sublayer, a conflict address adjudication sublayer, and an address-by-address projection sublayer; the hydraulic boundary gating sublayer generates a hydraulic boundary gating tensor based on the boundary mask tensor, the geometric boundary gating sublayer generates a geometric boundary gating tensor based on the boundary mask tensor, the conflict address adjudication sublayer generates a conflict adjudication flag based on the hydraulic boundary gating tensor, the geometric boundary gating tensor, and the loop constraint tensor, and the address-by-address projection sublayer generates boundary gating features based on the conflict adjudication flag.

[0017] Preferably, the address-by-address projection sublayer includes a node address projection channel, an edge address projection channel, and a loop address projection channel, where the node address, edge address, and loop address all belong to the shared topology address; the node address projection channel writes the boundary gating feature into the corresponding node address, the edge address projection channel writes the boundary gating feature into the corresponding edge address, and the loop address projection channel writes the boundary gating feature into the corresponding loop address.

[0018] Preferably, the loop-modulated cross-attention layer includes a co-address attention sub-layer, an adjacent attention sub-layer, a loop backhaul attention sub-layer, and a weight modulation input terminal; the weight modulation input terminal receives the loop constraint tensor and is respectively connected to the attention weight input terminals of the co-address attention sub-layer, the adjacent attention sub-layer, and the loop backhaul attention sub-layer.

[0019] Preferably, the same-address attention sublayer fuses boundary gating features and parameter control tensors based on the same shared topology address, the adjacent-edge attention sublayer fuses geometric flow node features and hydraulic flow node features corresponding to adjacent edges based on the node-edge routing table, and the loop backhaul attention sublayer backhauls the loop constraint tensor to the corresponding shared topology address based on the node-loop routing table, and generates a fused latent vector.

[0020] Preferably, the boundary recharge 3D decoding layer includes a parametric skeleton decoding sublayer, a surface patch decoding sublayer, a connection graph decoding sublayer, and a boundary recharge residual sublayer. The parametric skeleton decoding sublayer receives fused latent vectors and outputs a parametric skeleton tensor. The surface patch decoding sublayer receives fused latent vectors and boundary gated features and outputs a surface patch candidate tensor. The connection graph decoding sublayer receives fused latent vectors and boundary gated features and outputs a connection graph candidate tensor. The boundary recharge residual sublayer is connected to the surface patch decoding sublayer and the connection graph decoding sublayer.

[0021] Preferably, the boundary backfeeding 3D decoding layer further includes a topology candidate screening sublayer; the topology candidate screening sublayer receives parameter skeleton tensors, surface patch candidate tensors, connection graph candidate tensors, loop constraint tensors, and boundary gating features, and outputs a 3D parameterized geometry matrix and a topology assembly graph; the 3D parameterized geometry matrix includes a parameter control point matrix, a surface patch matrix, and a shared topology address marker matrix, and the topology assembly graph includes assembly nodes, assembly edges, and assembly loop markers.

[0022] In view of this, the present invention also provides a deep learning-driven three-dimensional parametric design system for water conservancy projects, including a shared topology address encoding module, an address hard routing dual-flow graph convolution module, a closed loop tensor constraint module, a boundary conflict gating projection module, a loop modulation cross attention module, and a boundary backfeed three-dimensional decoding module connected sequentially according to tensor data streams; The shared topology address encoding module is used to assign shared topology addresses to the three-dimensional parametric design data of water conservancy projects, and encode the three-dimensional parametric design data of water conservancy projects into node boundary tensors, edge adjacency tensors, parameter control tensors and boundary mask tensors carrying the same shared topology address, and generate node address embedding, node-edge routing table, node-parameter routing table and node-loop routing table; The address hard routing dual-flow graph convolution module is used to generate shared edge-gated matrices, geometric flow node features, and hydraulic flow node features; the closed loop tensor constraint module is used to generate loop constraint tensors; and the boundary conflict gated projection module is used to generate boundary gated features. The loop-modulated cross-attention module is used to modulate attention weights using loop constraint tensors and generate fused latent vectors. The boundary backfill 3D decoding module is used to generate 3D parametric geometry matrices and topology assembly diagrams.

[0023] Compared with existing technologies, the above technical solution has the following advantages: (i) This invention encodes the three-dimensional parametric design data of hydraulic engineering projects into node boundary tensors, edge adjacency tensors, parameter control tensors, and boundary mask tensors carrying the same shared topological address by setting a shared topological address encoding layer in the deep learning network topology. Furthermore, it generates node-edge routing tables, node-parameter routing tables, and node-loop routing tables, enabling forward propagation of geometric topology, hydraulic boundaries, parameter control, and loop relationships under a unified address system. This reduces index mismatches and feature misalignments during the propagation of data from different sources within the neural network, improving the consistency between the three-dimensional parametric geometric matrix and the topological assembly diagram.

[0024] (II) This invention generates geometric flow node features and hydraulic flow node features separately through a dual-flow graph convolutional layer using address hard routing. Furthermore, it controls the feature exchange between the geometric flow graph convolutional branches and the hydraulic flow graph convolutional branches via a shared edge gating matrix, preventing geometric and hydraulic features from being simply mixed and propagated within the same graph convolutional channel. This reduces the coupling interference between geometric constraint information and hydraulic boundary information, improving the feature representation stability of 3D parametric design data under multi-source constraints.

[0025] (III) This invention generates loop constraint tensors through a closed-loop tensor constraint layer, and uses these loop constraint tensors for attention weight modulation in subsequent loop modulation cross-attention layers. This allows closed topological relationships such as boundary loops, cross-sectional loops, and assembly loops to participate in the forward propagation process of the neural network. Therefore, it can reduce the probability of inconsistent surface patch connections, local topological breaks, and loop relationship mismatches during 3D geometry generation, and improve the closure consistency of the topological assembly diagram.

[0026] (iv) This invention generates boundary-gated features through a boundary conflict-gated projection layer, and performs conflict resolution and address-by-address projection on the hydraulic boundary-gated tensor and the geometric boundary-gated tensor under the same shared topological address, enabling the hydraulic boundary, geometric boundary, and structural connection boundary to form a structured gated propagation relationship at the intersection node. This reduces the impact of boundary constraint conflicts on the generation of 3D parametric geometry and improves the consistency of constraint propagation among boundary nodes, edge addresses, and loop addresses.

[0027] (V) This invention utilizes a boundary-recharged 3D decoding layer to integrate latent vectors and boundary-gated features for the generation of 3D parametric geometric matrices and topological assembly diagrams. A layered decoding structure is formed through a parametric skeleton decoding sublayer, a surface patch decoding sublayer, a connectivity graph decoding sublayer, and a boundary-recharged residual sublayer. This allows boundary constraint information to continue participating in the generation of surface patch candidate tensors and connectivity graph candidate tensors during the decoding stage, improving the coordination between the 3D parametric geometric output and the topological assembly relationship output. Attached Figure Description

[0028] Figure 1 A flowchart of a deep learning-driven three-dimensional parametric design method for hydraulic engineering. Figure 2 This is a diagram of the overall topology architecture of a deep learning network; Figure 3 This is a schematic diagram of the internal structure of a shared topology address coding layer; Figure 4 This is a schematic diagram of the internal structure of a dual-flow graph convolutional layer for address hard routing. Figure 5 This is a schematic diagram of the logic structure of a closed-loop tensor constraint layer and a boundary conflict gated projection layer. Figure 6 This is a schematic diagram of a tensor dimension transformation process; Figure 7 This is an abstract schematic diagram of a three-dimensional parametric diagram structure for a water conservancy project. Figure 8 This is a schematic diagram showing the performance comparison of a model and the results of an ablation experiment. Detailed Implementation

[0029] The following description, in conjunction with the accompanying drawings, details a deep learning-driven three-dimensional parametric design method and system for hydraulic engineering. The embodiments described below are for illustrating specific implementations of this technical solution and should not be construed as the sole limitation on its implementation. Without departing from the deep learning network topology and its data flow connections, those skilled in the art can adaptively adjust the number of neurons, feature dimensions, activation functions, training batch size, data format, and deployment hardware for some network layers.

[0030] like Figure 1 and Figure 2 As shown, in this embodiment, the deep learning network topology includes a shared topology address encoding layer, an address hard routing dual-flow graph convolutional layer, a closed-loop tensor constraint layer, a boundary conflict gated projection layer, a loop modulation cross-attention layer, and a boundary backfeed 3D decoding layer, all connected sequentially. This deep learning network topology is used for tensor processing and network forward propagation of 3D parametric design data for hydraulic engineering projects. Instead of directly inputting hydraulic engineering parameters as ordinary input data into a general neural network, it establishes data flow connections between the shared topology address, address hard routing, dual-flow graph convolution, closed-loop tensor constraint, boundary conflict gated projection, loop modulation cross-attention, and boundary backfeed 3D decoding layers within the neural network.

[0031] In this embodiment, the three-dimensional parametric design data for hydraulic engineering projects can originate from B-rep boundary representation data of hydraulic structures, IFC format model data, parametric surface control point data, design head data, spillway boundary data, structural zoning marker data, dam or gate chamber outline data, and connecting component constraint data. Taking the three-dimensional parametric model of a spillway as an example, the original data can include the gate pier outline curve, bottom plate boundary surface, gate opening boundary line, stilling basin boundary surface, design head, allowable flow velocity, material zoning, and component assembly relationships. After preprocessing, the above original data forms a topological representation of nodes, edges, loops, and parametric control points, such as... Figure 7 As shown.

[0032] In the data preprocessing stage, node sets, edge sets, loop sets, and parameter control point sets are first extracted from the 3D parametric design data of hydraulic engineering projects. Node sets correspond to geometric curve endpoints, surface sampling points, boundary control points, and hydraulic boundary sampling points; edge sets correspond to the connection relationships between adjacent nodes; loop sets correspond to boundary loops, cross-sectional loops, or assembly loops formed by multiple edges; and parameter control point sets correspond to parameter points of control curves, control cross-sections, surface patches, and assembly connection relationships. Subsequently, shared topology addresses are assigned to nodes, edges, loops, and parameter control points, enabling address alignment of data from different sources within the same deep learning network topology.

[0033] In one implementation, the number of nodes is denoted as... The number of edges is denoted as The number of parameter control points is denoted as The number of ring roads is recorded as The node boundary tensor is denoted as... The edge-adjacent tensor is denoted as The parameter control tensor is denoted as The boundary mask tensor is denoted as The node boundary tensor represents node coordinates, node normals, node curvature, the structural partition to which the node belongs, and node boundary attributes. The edge adjacency tensor represents the start and end addresses, edge type, edge length, and adjacent surface patch indices of edges. The parameter control tensor represents curve control points, section control points, surface patch control points, and hydraulic parameter control values. The boundary mask tensor indicates whether a node belongs to an inlet boundary, outlet boundary, flow continuity boundary, geometric profile boundary, datum plane boundary, or structural partition boundary. The normalized tensor features are dimensionless, while the unnormalized engineering physical quantities retain their original units.

[0034] In one specific example, the spillway model was obtained after sampling. Each node Strip, Each parameter control point and There are several loops. The dimension of the node boundary tensor is:

[0035] The dimension of the edge adjacency tensor is:

[0036] The dimension of the parameter-controlled tensor is:

[0037] The dimension of the boundary mask tensor is:

[0038] in, Represents the node boundary tensor. This represents the edge-adjacency tensor. This indicates a parameter control tensor. Represents the boundary mask tensor. Represents the real number field. This represents the set of Boolean values; 1248 represents the number of nodes; 3580 represents the number of edges; 412 represents the number of parameter control points; 32 represents the feature dimension of the node boundary tensor; 16 represents the feature dimension of the edge adjacency tensor; and 24 represents the feature dimension of the parameter control tensor. This represents the number of boundary categories in the boundary mask tensor. The number of nodes, edges, parameter control points, and loops mentioned above are only examples of data scale in one embodiment, and can be adjusted according to the model complexity in other three-dimensional parametric models of hydraulic engineering.

[0039] For hydraulic parameters, a normalization method can be used to write them into the parameter control tensor. Taking the design head as an example, let the original design head value be... The lower limit of the design head is The upper limit of the design head is Normalized design head is Then we have:

[0040] in, This represents the original design head value, in meters (m). This indicates the lower limit of the design head, in meters (m). This indicates the upper limit of the design head, in meters (m). This represents the normalized design head, a dimensionless value. If the design head of a certain spillway is 78.5m, the lower limit of the design head is 40.0m, and the upper limit of the design head is 120.0m, then the normalized design head is:

[0041] The normalized design head is written into the parameter control tensor through the parameter control point channel corresponding to the hydraulic boundary node. For parameters such as velocity limits, gate width, bottom elevation, and boundary elevation difference, normalization can also be used to write them into the parameter control tensor, with different physical quantities occupying different parameter channels.

[0042] like Figure 3 As shown, the shared topology address includes node address, edge address, loop address, and parameter control address. The node boundary tensor carries the node address, the edge adjacency tensor carries the edge address, the parameter control tensor carries the parameter control address, and the boundary mask tensor carries the node address and loop address. The shared topology address encoding layer includes an address allocation sublayer, an address alignment sublayer, an address routing table generation sublayer, and a suspended address isolation sublayer. The address allocation sublayer assigns shared topology addresses to nodes, edges, loops, and parameter control points, respectively. The address alignment sublayer aligns node boundary tensors, boundary mask tensors, and parameter control tensors with the same node address. The address routing table generation sublayer generates node-edge routing tables, node-parameter routing tables, and node-loop routing tables. The suspended address isolation sublayer writes shared topology addresses that do not form routing connections in the node-edge routing table, node-parameter routing table, or node-loop routing table into suspended address channels. The output of the suspended address channels is connected to the address hard routing dual-flow graph convolutional layer.

[0043] In this embodiment, the suspended address can be an isolated parameter control point, an endpoint of an unclosed boundary curve, a temporary node lacking a corresponding edge relationship, or a boundary sampling point that has not yet participated in the formation of a loop. After the suspended address isolation sublayer writes this type of shared topology address into the suspended address channel, the subsequent address hard routing dual-flow graph convolutional layer can process it as a bypass input, allowing the main graph convolutional path to prioritize processing nodes and edges that have formed clear routing relationships. The suspended address channel can be implemented using sparse tensor caching, independent address queues, zero-padding masks, missing address flags, or bypass residual connections.

[0044] In the shared topology address coding layer, node address embedding can be implemented using an embedding matrix. Let the shared topology address coding matrix be... The address embedding weight matrix is The address embedding bias vector is The node address is embedded as Then we have:

[0045] in, Represents the shared topology address encoding matrix. This represents the address embedding weight matrix. This represents the address embedding bias vector. This represents the node address embedding. The address embedding weight matrix and address embedding bias vector can be obtained through training, or they can be randomly generated during the model initialization phase and then used for backpropagation updates.

[0046] like Figure 4 As shown, the address hard-routing dual-flow graph convolutional layer includes a geometric flow graph convolutional branch, a hydraulic flow graph convolutional branch, a shared-edge gating sublayer, and an interflow residual exchange sublayer. The geometric flow graph convolutional branch processes geometric information such as node coordinates, surface normals, curvature, edge lengths, and surface patch indices; the hydraulic flow graph convolutional branch processes hydraulic information such as design head, flow direction continuity markers, inlet boundary markers, outlet boundary markers, and velocity limits. The shared-edge gating sublayer generates a shared-edge gating matrix based on the node-edge routing table, the node-parameter routing table, and the boundary mask tensor. Both the geometric flow graph convolutional branch and the hydraulic flow graph convolutional branch receive the shared-edge gating matrix. The interflow residual exchange sublayer includes a geometry-to-hydraulic exchange channel and a hydraulic-to-geometric exchange channel. The geometry-to-hydraulic exchange channel writes the geometric flow node features into the hydraulic flow graph convolutional branch according to the shared edge gating matrix, and the hydraulic-to-geometry exchange channel writes the hydraulic flow node features into the geometric flow graph convolutional branch according to the shared edge gating matrix.

[0047] In one implementation, let the first... The geometric flow node characteristics of the layer are , No. The hydraulic flow node characteristics of the layer are The normalized adjacency matrix is The shared side-gating matrix is , No. The layer geometric flow self-updating weight matrix is , No. The self-updating weight matrix of the hydraulic flow is , No. The layer geometry to hydraulic exchange weight matrix is , No. The hydraulic-to-geometric exchange weight matrix is , No. Layer geometric flow bias vector is , No. The hydraulic flow bias vector is The activation function is The geometric flow node feature update process is represented as follows:

[0048] The process of updating the hydraulic flow node features is represented as follows:

[0049] in, Indicates the first Geometric flow node characteristics of the layer Indicates the first Hydraulic flow node characteristics of the layer, Indicates the first Geometric flow node characteristics of the layer Indicates the first Hydraulic flow node characteristics of the layer, Let represent the normalized adjacency matrix constructed from the edge adjacency tensor. Indicates a shared side-gated matrix. Indicates the first Layer geometry flow self-updating weight matrix, Indicates the first Layer hydraulic flow self-updating weight matrix, Indicates the first Layer geometry to hydraulic exchange weight matrix, Indicates the first Hydraulic-to-geometric exchange weight matrix Indicates the first Layer geometry flow bias vector Indicates the first Layer hydraulic flow bias vector This indicates element-wise multiplication. This represents the linear rectification activation function.

[0050] In this embodiment, geometric flow node features and hydraulic flow node features are not directly mixed in the same channel, but rather undergo cross-flow residual exchange under the constraint of a shared edge gating matrix. The shared edge gating matrix can be generated using node-edge routing tables and node-parameter routing tables, and its numerical range can be constrained by a sigmoid activation function. to Between. The closer the shared side-gated matrix is. This indicates that strong cross-flow feature exchange is allowed between the corresponding nodes; the closer the shared side gating matrix is ​​to the desired value... This indicates that cross-flow feature exchange between corresponding nodes is suppressed. The shared edge gating sublayer can be implemented using fully connected layers, sparse linear layers, or graph attention gating layers.

[0051] like Figure 5As shown, the closed-loop tensor constraint layer includes a boundary loop extraction sublayer, a cross-sectional loop extraction sublayer, an assembly loop extraction sublayer, a loop direction encoding sublayer, and a loop tensor generation sublayer. The boundary loop extraction sublayer extracts boundary loop tensors based on the node-loop routing table. The cross-sectional loop extraction sublayer extracts cross-sectional loop tensors based on edge adjacency tensors. The assembly loop extraction sublayer extracts assembly loop tensors based on parameter control tensors. The loop direction encoding sublayer generates loop direction tensors based on the order of shared topology addresses corresponding to adjacent edges in the same loop and writes the loop direction tensors into the loop constraint tensors. The loop tensor generation sublayer generates loop constraint tensors based on the boundary loop tensors, cross-sectional loop tensors, assembly loop tensors, and loop direction tensors.

[0052] Let the boundary loop tensor be The cross-sectional loop tensor is The assembly loop tensor is The loop direction tensor is The loop fusion weight matrix is The loop fusion bias vector is The loop constraint tensor is The loop constraint tensor is then expressed as:

[0053] in, Represents the boundary loop tensor. Represents the cross-sectional loop tensor. This represents the assembly loop tensor. Represents the loop direction tensor. This represents the loop fusion weight matrix. This represents the loop fusion bias vector. This represents the loop constraint tensor. This represents a tensor splicing operation. In another implementation, the loop tensor generation sublayer can also generate loop constraint tensors using element-wise addition, gated weighted summation, or multilayer perceptron fusion.

[0054] The boundary conflict gating projection layer comprises a hydraulic boundary gating sublayer, a geometric boundary gating sublayer, a conflict address adjudication sublayer, and a per-address projection sublayer. The hydraulic boundary gating sublayer generates a hydraulic boundary gating tensor based on the boundary mask tensor. The geometric boundary gating sublayer generates a geometric boundary gating tensor based on the boundary mask tensor. The conflict address adjudication sublayer generates conflict adjudication flags based on the hydraulic boundary gating tensor, the geometric boundary gating tensor, and the loop constraint tensor. The per-address projection sublayer includes node address projection channels, edge address projection channels, and loop address projection channels. Node addresses, edge addresses, and loop addresses all belong to the shared topology address. The node address projection channel writes the boundary gating features into the corresponding node address; the edge address projection channel writes the boundary gating features into the corresponding edge address; and the loop address projection channel writes the boundary gating features into the corresponding loop address.

[0055] Let the hydraulic boundary gate tensor be... The geometric boundary gate tensor is Conflict ruling marked as Boundary gating features are The hydraulic gating weight matrix is The geometric gate weight matrix is The loop gating weight matrix is The boundary gate bias vector is The activation function is The hydraulic boundary gate tensor and the geometric boundary gate tensor are then expressed as:

[0056] The conflict resolution marker is represented as:

[0057] Boundary gating features are represented as follows:

[0058] in, Represents the hydraulic boundary gating tensor. Represents the geometric boundary gate tensor. Indicates conflict resolution mark, Indicates boundary gating features, This represents the hydraulic gating weight matrix. Represents the geometrically gated weight matrix. This represents the loop gating weight matrix. This represents the boundary gating bias vector. This represents the Sigmoid activation function. This indicates element-wise multiplication. The above expression represents one feasible gating method. In another implementation, the conflict resolution label can also be generated by gating neurons, multilayer perceptrons, or sparse attention gating units.

[0059] like Figure 2 and Figure 6 As shown, the loop-modulated cross-attention layer is used to modulate attention weights using the loop constraint tensor and generate a fused latent vector. The loop-modulated cross-attention layer includes a co-address attention sublayer, a neighboring edge attention sublayer, a loop backhaul attention sublayer, and a weight modulation input. The weight modulation input receives the loop constraint tensor and is connected to the attention weight inputs of the co-address attention sublayer, the neighboring edge attention sublayer, and the loop backhaul attention sublayer, respectively. The co-address attention sublayer fuses boundary gating features and parameter control tensors based on the same shared topology address. The neighboring edge attention sublayer fuses the geometric flow node features and hydraulic flow node features corresponding to adjacent edges based on the node-edge routing table. The loop backhaul attention sublayer backhauls the loop constraint tensor to the corresponding shared topology address based on the node-loop routing table and generates a fused latent vector.

[0060] Let the query matrix be The key matrix is The value matrix is The loop modulation matrix is Attention dimension is The attention score matrix is The fused latent vectors are The query matrix is ​​obtained by linear mapping of boundary gating features, the key and value matrices are obtained by linear mapping of parameter control tensors, geometric flow node features, and hydraulic flow node features, and the loop modulation matrix is ​​obtained by linear mapping of loop constraint tensors. The attention score matrix is ​​represented as:

[0061] The fused latent vector is represented as:

[0062] in, Represents the query matrix. Represents the key matrix, Represents a value matrix, Represents the loop modulation matrix. Represents the attention dimension. Represents the attention score matrix. This represents the fused latent vector. The transpose of the key matrix. This represents the normalized exponential function. The loop modulation matrix participates in the calculation of the attention score matrix, enabling the loop constraint tensor to participate in forward propagation during the attention weight calculation stage.

[0063] The boundary-feedback 3D decoding layer is used to generate a 3D parametric geometry matrix and a topology assembly diagram. This layer includes a parametric skeleton decoding sublayer, a surface patch decoding sublayer, a connectivity graph decoding sublayer, a boundary-feedback residual sublayer, and a topology candidate filtering sublayer. The parametric skeleton decoding sublayer receives fused latent vectors and outputs a parametric skeleton tensor. The surface patch decoding sublayer receives fused latent vectors and boundary-gated features and outputs a surface patch candidate tensor. The connectivity graph decoding sublayer receives fused latent vectors and boundary-gated features and outputs a connectivity graph candidate tensor. The boundary-feedback residual sublayer connects to the surface patch decoding sublayer and the connectivity graph decoding sublayer. The topology candidate filtering sublayer receives the parametric skeleton tensor, surface patch candidate tensor, connectivity graph candidate tensor, loop constraint tensor, and boundary-gated features, and outputs a 3D parametric geometry matrix and a topology assembly diagram. The 3D parametric geometry matrix includes a parametric control point matrix, a surface patch matrix, and a shared topology address marker matrix. The topology assembly diagram includes assembly nodes, assembly edges, and assembly loop markers.

[0064] Let the parametric skeleton tensor be... The candidate tensor of the curved surface patch is The candidate tensor of the connection graph is The three-dimensional parametric geometry matrix is The topology assembly diagram is as follows The parameter skeleton weight matrix is The surface patch decoding weight matrix is The connection graph decoding weight matrix is The surface patch boundary backfill weight matrix is The boundary backfill weight matrix of the connection graph is The surface patch decoding bias vector is The connection graph decoding bias vector is The parametric skeleton tensor is represented as:

[0065] The candidate tensor of the surface patch is represented as:

[0066] The candidate tensor of the connected graph is represented as:

[0067] in, Represents the parametric skeleton tensor. Represents the candidate tensor of the surface patch. This represents a candidate tensor for the connectivity graph. Represents a three-dimensional parametric geometry matrix. Represents the topology assembly diagram. This represents the parameter skeleton weight matrix. This represents the surface patch decoding weight matrix. This represents the weight matrix of the connection graph decoding. This represents the weight matrix for the backfilling of the surface patch boundary. This represents the boundary backfeed weight matrix of the connectivity graph. This represents the surface patch decoding bias vector. This represents the bias vector for connection graph decoding. Boundary refeedback residual sublayers enable boundary-gated features to enter the generation process of surface patch candidate tensors and connection graph candidate tensors during the decoding stage.

[0068] The topology candidate selection sublayer calculates candidate selection scores based on the parametric skeleton tensor, surface patch candidate tensor, connectivity graph candidate tensor, loop constraint tensor, and boundary gating features. Let the topology candidate selection score be... The parameter skeleton candidate screening weight matrix is ​​as follows: The candidate selection weight matrix for curved surface patches is: The candidate selection weight matrix of the connection graph is: The candidate selection weight matrix with loop constraints is: The boundary-gated candidate selection weight matrix is ​​as follows: The candidate selection bias vector is Then we have:

[0069] in, Indicates the topology candidate selection score. This represents the candidate selection weight matrix of the parameter skeleton. This represents the candidate selection weight matrix for curved surface patches. This represents the candidate selection weight matrix of the connection graph. This represents the candidate selection weight matrix for loop constraints. This represents the boundary-gated candidate selection weight matrix. This represents the candidate selection bias vector. The topology candidate selection sublayer selects surface patch connection relationships and assembly connection relationships based on the topology candidate selection score, and generates a three-dimensional parametric geometry matrix and a topology assembly diagram.

[0070] In this embodiment, the model training process can be performed using supervised training. The training method is used to illustrate the parameter learning process of the deep learning network topology and does not constitute a restriction on the forward propagation network topology. The training data can use a self-built hydraulic structure B-rep topology dataset. This dataset can be generated from B-rep models, IFC models, and parametric modeling scripts, and the correspondence between nodes, edges, loops, and parameter control points can be confirmed manually or through rule-based verification. In a set of simulation validation data, the dataset includes 8000 parametric hydraulic structure models, of which 6000 are used as the training set, 1000 as the validation set, and 1000 as the test set. Each model contains node boundary tensors, edge adjacency tensors, parameter control tensors, boundary mask tensors, target 3D parametric geometry matrices, and target topology assembly diagrams.

[0071] During training, the Adam optimizer can be used, with the learning rate set to... to The batch size can be set to 8 to 32, and the number of training rounds can be set to 100 to 300. Alternatively, a learning rate cosine annealing strategy or a segmented decay strategy can be used.

[0072] To enable the network to simultaneously learn geometry generation, topology matching, loop closure, boundary constraints, and address consistency, this embodiment uses a total loss function for training. Let the total loss function be... The geometric reconstruction loss is The topology matching loss is The loop closure loss is Boundary constraint loss is Address consistency loss is The geometric loss weights are The topological loss weight is The loop loss weight is The boundary loss weights are The address loss weight is Then we have:

[0073] in, Represents the total loss function. Represents the geometric reconstruction loss. Represents the topology matching loss. Indicates the loss due to loop closure. Represents the boundary constraint loss. This indicates address consistency loss. Represents the geometric loss weights. Represents the topological loss weights. Indicates the loop loss weight, Indicates the boundary loss weights. This represents the address loss weight. In one implementation, the geometric loss weight is set to 1.00, the topology loss weight to 0.80, the loop loss weight to 1.20, the boundary loss weight to 1.00, and the address loss weight to 0.50. Each loss weight can be adjusted based on the validation set error.

[0074] Geometric reconstruction loss can be calculated using the average distance error between the sampling points of the output surface patch and the sampling points of the target surface patch; topology matching loss can be calculated using the edge connection cross-entropy between the predicted topology assembly and the target topology assembly; loop closure loss can be calculated using the offset of the start and end nodes of the predicted loop and the loop edge sequence error; boundary constraint loss can be calculated using the boundary node out-of-bounds ratio and the boundary mask classification error; address consistency loss can be calculated using the matching error between the predicted shared topology address and the target shared topology address.

[0075] In a series of simulations, the deep learning network topology of this embodiment is compared with the UV-Net representation-based deep learning processing scheme for 3D CAD objects (which is closest to existing technologies), the standard GCN network, the PointNet++ network, and the general Transformer network. The comparative experiments use the same training data, the same test data, and the same number of training rounds. All models input 3D parametric design data for hydraulic engineering from the same source and output 3D parametric geometric matrices and topology assembly diagrams. Evaluation metrics include topology consistency score, loop closure error, boundary out-of-bounds rate, and parameter control point mismatch rate. A higher topology consistency score indicates a more consistent topology assembly relationship; a lower loop closure error indicates a more stable closed loop; a lower boundary out-of-bounds rate indicates a higher degree of boundary constraint satisfaction; and a lower parameter control point mismatch rate indicates a more accurate correspondence between output parameter control points and target parameter control points.

[0076] Table 1. Performance Comparison Results of Different Models

[0077] As shown in Table 1, under the simulation verification conditions, the deep learning network topology in this embodiment shows numerical improvements in topology consistency score, loop closure error, boundary out-of-bounds rate, and parameter control point mismatch rate. Specifically, the loop closure error is improved by the UV-Net representation-based deep learning processing scheme for 3D CAD objects. Change to The boundary out-of-bounds rate changed from 8.70% to 2.10%. The above data is used to illustrate the structural constraint effect of shared topology address, closed loop tensor constraint, boundary conflict gated projection, and boundary refeeding 3D decoding on the generation of 3D parametric geometry during forward propagation.

[0078] To further verify the impact of each network layer structure on the overall output, ablation experiments were conducted under the same dataset and training conditions. The ablation experiments removed the suspended address isolation sublayer in the shared topology address coding layer, the cross-flow residual exchange sublayer in the address hard routing dual-flow graph convolutional layer, the closed-loop tensor constraint layer, the boundary conflict gated projection layer, the weighted modulation input in the loop modulation cross-attention layer, and the boundary refeed residual sublayer in the boundary refeed 3D decoding layer.

[0079] Table 2 Ablation Experiment Results

[0080] As shown in Table 2, under the simulation verification conditions, after removing the closed loop tensor constraint layer, the loop closure error is reduced from... Change to After removing the boundary conflict-gated projection layer, the boundary out-of-bounds rate changed from 2.10% to 8.80%; after removing the cross-flow residual exchange sublayer, the topology consistency score changed from 98.50% to 94.70%. The above ablation experiment data illustrate the hierarchical cooperation relationship between shared topology address coding, address hard routing dual-flow graph convolution, closed-loop tensor constraints, boundary conflict-gated projection, loop modulation cross-attention, and boundary refeeding 3D decoding.

[0081] In terms of system implementation, the shared topology address coding layer can be implemented as a shared topology address coding module, the address hard routing dual-flow graph convolutional layer can be implemented as an address hard routing dual-flow graph convolutional module, the closed loop tensor constraint layer can be implemented as a closed loop tensor constraint module, the boundary conflict gating projection layer can be implemented as a boundary conflict gating projection module, the loop modulation cross-attention layer can be implemented as a loop modulation cross-attention module, and the boundary backfeed 3D decoding layer can be implemented as a boundary backfeed 3D decoding module. Each module is implemented according to... Figure 2 The tensor data streams shown are connected sequentially. The above modules can be deployed on a graphics processor, tensor processor, or general-purpose computing processor, and the node-edge routing table, node-parameter routing table, and node-loop routing table can be queried through a sparse tensor computing framework.

[0082] In one implementation, a shared topology address encoding module receives 3D parametric design data from a hydraulic engineering project and encodes it into node boundary tensors, edge adjacency tensors, parameter control tensors, and boundary mask tensors. Simultaneously, it generates node address embeddings, node-edge routing tables, node-parameter routing tables, and node-loop routing tables. An address hard-routing dual-flow graph convolution module generates a shared edge-gated matrix, geometric flow node features, and hydraulic flow node features. A closed-loop tensor constraint module generates loop constraint tensors. A boundary conflict-gated projection module generates boundary-gated features. A loop-modulated cross-attention module modulates attention weights using the loop constraint tensors and generates fused latent vectors. A boundary-backfeed 3D decoding module generates a 3D parametric geometric matrix and a topology assembly diagram.

[0083] In the above implementation, the focus of the deep learning network topology is on the hierarchical connections and forward propagation data flow within the artificial neural network. A shared topology address encoding layer, an address hard-routing dual-flow graph convolutional layer, a closed-loop tensor constraint layer, a boundary conflict-gated projection layer, a loop modulation cross-attention layer, and a boundary-feedback 3D decoding layer together constitute the deep learning network topology structure for 3D parametric design data of water conservancy projects.

[0084] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A deep learning-driven three-dimensional parametric design method for hydraulic engineering, characterized in that, Constructing and running a deep learning network topology, wherein the deep learning network topology comprises, in sequence, a shared topology address encoding layer, an address hard-routing dual-flow graph convolutional layer, a closed-loop tensor constraint layer, a boundary conflict-gated projection layer, a loop modulation cross-attention layer, and a boundary refeeding 3D decoding layer; the method includes: A shared topology address is assigned to the three-dimensional parametric design data of the water conservancy project, and the three-dimensional parametric design data of the water conservancy project is encoded into a node boundary tensor, an edge adjacency tensor, a parameter control tensor and a boundary mask tensor carrying the same shared topology address; The shared topology address coding layer generates node address embeddings, node-edge routing tables, node-parameter routing tables, and node-loop routing tables. The address hard routing dual-flow graph convolutional layer generates a shared edge gating matrix, geometric flow node features, and hydraulic flow node features; The closed-loop tensor constraint layer generates a loop constraint tensor; Boundary gated features are generated by the boundary conflict gating projection layer; The loop-modulated cross-attention layer uses the loop constraint tensor to modulate attention weights and generate a fused latent vector; The three-dimensional parametric geometry matrix and topology assembly diagram are generated by the boundary backfilling three-dimensional decoding layer.

2. The deep learning-driven three-dimensional parametric design method for hydraulic engineering according to claim 1, characterized in that, The shared topology address includes node address, edge address, loop address, and parameter control address; the node boundary tensor carries the node address, the edge adjacency tensor carries the edge address, the parameter control tensor carries the parameter control address, and the boundary mask tensor carries the node address and the loop address.

3. The deep learning-driven three-dimensional parametric design method for hydraulic engineering according to claim 2, characterized in that, The shared topology address encoding layer includes an address allocation sublayer, an address alignment sublayer, and an address routing table generation sublayer. The address allocation sublayer allocates the shared topology addresses to nodes, edges, loops, and parameter control points, respectively. The address alignment sublayer aligns the node boundary tensors, boundary mask tensors, and parameter control tensors with the same node addresses. The address routing table generation sublayer generates the node-edge routing table, the node-parameter routing table, and the node-loop routing table.

4. The deep learning-driven three-dimensional parametric design method for hydraulic engineering according to claim 3, characterized in that, The shared topology address encoding layer further includes a suspended address isolation sublayer; the suspended address isolation sublayer writes the shared topology address that does not form a routing connection in the node-edge routing table, the node-parameter routing table, or the node-loop routing table into the suspended address channel, where the shared topology address is a node address, edge address, loop address, or parameter control address, and the output of the suspended address channel is connected to the address hard routing dual-flow graph convolutional layer.

5. The deep learning-driven three-dimensional parametric design method for hydraulic engineering according to claim 1, characterized in that, The address hard routing dual-flow graph convolutional layer includes a geometric flow graph convolutional branch, a hydraulic flow graph convolutional branch, and a shared edge gating sublayer; the shared edge gating sublayer generates the shared edge gating matrix based on the node-edge routing table, the node-parameter routing table, and the boundary mask tensor, and both the geometric flow graph convolutional branch and the hydraulic flow graph convolutional branch receive the shared edge gating matrix.

6. The deep learning-driven three-dimensional parametric design method for hydraulic engineering according to claim 5, characterized in that, The address hard routing dual-flow graph convolutional layer further includes a cross-flow residual switching sublayer; the cross-flow residual switching sublayer includes a geometry-to-hydraulic switching channel and a hydraulic-to-geometry switching channel, wherein the geometry-to-hydraulic switching channel writes the geometric flow node features into the hydraulic flow graph convolutional branch according to the shared edge gating matrix, and the hydraulic-to-geometry switching channel writes the hydraulic flow node features into the geometric flow graph convolutional branch according to the shared edge gating matrix.

7. The deep learning-driven three-dimensional parametric design method for hydraulic engineering according to claim 1, characterized in that, The closed loop tensor constraint layer includes a boundary loop extraction sublayer, a cross-sectional loop extraction sublayer, an assembly loop extraction sublayer, and a loop tensor generation sublayer. The boundary loop extraction sublayer extracts boundary loop tensors based on the node-loop routing table. The cross-sectional loop extraction sublayer extracts cross-sectional loop tensors based on the edge adjacency tensors. The assembly loop extraction sublayer extracts assembly loop tensors based on the parameter control tensors. The loop tensor generation sublayer generates the loop constraint tensors based on the boundary loop tensors, the cross-sectional loop tensors, and the assembly loop tensors.

8. The deep learning-driven three-dimensional parametric design method for hydraulic engineering according to claim 7, characterized in that, The closed loop tensor constraint layer further includes a loop direction encoding sublayer; the loop direction encoding sublayer generates a loop direction tensor according to the arrangement order of the shared topology addresses corresponding to adjacent edges in the same loop, and writes the loop direction tensor into the loop constraint tensor.

9. The deep learning-driven three-dimensional parametric design method for hydraulic engineering according to claim 1, characterized in that, The boundary conflict gating projection layer includes a hydraulic boundary gating sublayer, a geometric boundary gating sublayer, a conflict address adjudication sublayer, and a per-address projection sublayer. The hydraulic boundary gating sublayer generates a hydraulic boundary gating tensor based on the boundary mask tensor. The geometric boundary gating sublayer generates a geometric boundary gating tensor based on the boundary mask tensor. The conflict address adjudication sublayer generates a conflict adjudication flag based on the hydraulic boundary gating tensor, the geometric boundary gating tensor, and the loop constraint tensor. The per-address projection sublayer generates the boundary gating feature based on the conflict adjudication flag.

10. A deep learning-driven three-dimensional parametric design system for hydraulic engineering, characterized in that, It includes a shared topology address encoding module that connects tensor data streams sequentially, an address hard routing dual-stream graph convolution module, a closed loop tensor constraint module, a boundary conflict gated projection module, a loop modulation cross-attention module, and a boundary backfeed 3D decoding module; The shared topology address encoding module is used to allocate shared topology addresses to the three-dimensional parametric design data of water conservancy projects, and to encode the three-dimensional parametric design data of water conservancy projects into node boundary tensors, edge adjacency tensors, parameter control tensors and boundary mask tensors carrying the same shared topology address, and to generate node address embedding, node-edge routing table, node-parameter routing table and node-loop routing table; The address hard routing dual-flow graph convolution module is used to generate shared edge gating matrices, geometric flow node features, and hydraulic flow node features; the closed loop tensor constraint module is used to generate loop constraint tensors; and the boundary conflict gating projection module is used to generate boundary gating features. The loop-modulated cross-attention module is used to modulate attention weights using the loop constraint tensor and generate a fused latent vector. The boundary backfilling 3D decoding module is used to generate 3D parametric geometric matrices and topological assembly diagrams.

Images