Deep learning-based building design scheme intelligent generation method and system

Abstract

The application discloses a kind of based on deep learning's building design scheme intelligent generation method and system, comprising: obtaining building design demand and a plurality of contrast building schemes based on consistent comparison design paradigm;Extract the semantic feature of demand and the comparison design paradigm feature of each contrast scheme;Integrate the paradigm feature of multiple contrast schemes and obtain the second integrated design paradigm feature by graph convolution network coding;With the paradigm feature as generation constraint, generate the selected building scheme that meets design demand based on semantic feature of demand.The application improves the rationality and practicability of generated scheme by fusing the consistent paradigm constraint of multiple contrast schemes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for intelligent generation of architectural design schemes based on deep learning. Background Technology

[0002] Current architectural design scheme generation mainly relies on the human experience of designers, requiring repeated comparisons and adjustments, which is inefficient and easily affected by subjective factors. Some deep learning-based methods only focus on the semantic features of a single requirement and do not combine the consistent design paradigm of multiple architectural schemes for comparison, resulting in a large deviation between the generated schemes and the industry's common paradigms, making it difficult to meet the paradigm constraints of actual projects. There is an urgent need for an intelligent generation method that integrates the paradigm constraints of multiple schemes for comparison. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligent generation of architectural design schemes based on deep learning.

[0004] In a first aspect, embodiments of the present invention provide a method for intelligent generation of architectural design schemes based on deep learning, comprising: Obtain architectural design requirements and multiple comparative architectural schemes, the multiple comparative architectural schemes being based on a consistent comparative design paradigm, and the architectural design requirements being used to describe the design requirements of the architectural design schemes. Obtain the semantic features of the architectural design requirements, and obtain the comparative design paradigm features of each comparative architectural scheme in the multiple comparative architectural schemes. By integrating the comparative design paradigm features corresponding to the multiple comparative architectural schemes, the first integrated design paradigm feature is obtained; The first integrated design paradigm feature is processed by graph convolutional network encoding to obtain the second integrated design paradigm feature, which is used to describe the comparative design paradigm. Using the comparative design paradigm of the second integrated design paradigm features as the generation constraint, candidate architectural schemes are generated based on the semantic features of the requirements. The architectural design of the candidate architectural schemes meets the design requirements defined by the architectural design requirements.

[0005] In one possible implementation, the method is performed based on a first architectural design generation model, which includes a feature extraction module and a first scale compression module; obtaining the comparative design paradigm features corresponding to each of the multiple comparative architectural schemes includes: Based on the feature extraction module, feature extraction processing is performed on each of the comparative architectural schemes to obtain the architectural visual coding features corresponding to each of the comparative architectural schemes. Based on the first scale compression module, feature compression is performed on the architectural visual coding features corresponding to each of the comparative architectural schemes to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes.

[0006] In one possible implementation, the first-scale compression module includes a first visual dimensionality reduction mapping component and a graph convolution processing component, wherein the graph convolution processing component includes a first spatial aggregation component and a second spatial aggregation component; the step of compressing the architectural visual encoding features corresponding to each of the comparative architectural schemes based on the first-scale compression module to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes includes: Based on the first visual dimensionality reduction mapping component, the architectural visual coding features corresponding to each of the comparative architectural schemes are converted into the first visual dimensionality reduction features corresponding to each of the comparative architectural schemes. Based on the first spatial aggregation component, graph convolutional network encoding is performed on the first visual dimensionality reduction features corresponding to each of the comparative architectural schemes to obtain the first transition features corresponding to each of the comparative architectural schemes. Based on the second spatial aggregation component, graph convolutional network encoding is performed on the first transition features corresponding to each of the comparative architectural schemes to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes.

[0007] In one possible implementation, the first-scale compression module includes a plurality of sequentially cascaded first feature compression components; the step of compressing the architectural visual coding features corresponding to each of the comparative architectural schemes based on the first-scale compression module to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes includes: Based on the first first feature compression component among the plurality of first feature compression components, feature compression is performed on the architectural visual coding features corresponding to each of the comparative architectural schemes to obtain the first comparative design paradigm feature corresponding to each of the comparative architectural schemes. Based on the target first feature compression component among the plurality of first feature compression components, feature compression is performed on the preceding comparative design paradigm features corresponding to each of the comparative architectural schemes to obtain the target comparative design paradigm features corresponding to each of the comparative architectural schemes, wherein the preceding comparative design paradigm feature is the one preceding the target first feature compression component.

[0008] In one possible implementation, the first architectural design generation model includes a feature integration module, which includes multiple feature integration components; the integration of the comparative design paradigm features corresponding to the multiple comparative architectural schemes to obtain the first integrated design paradigm feature includes: Based on the first feature integration component among the multiple feature integration components, the first comparative design paradigm features corresponding to each of the multiple comparative architectural schemes are integrated to obtain the first integrated design paradigm feature; Based on the target feature integration component among the multiple feature integration components, the target comparative design paradigm features corresponding to each of the multiple comparative architectural schemes are integrated to obtain the target first integrated design paradigm features.

[0009] In one possible implementation, the feature integration module further includes multiple third-space aggregation components; the step of performing graph convolutional network encoding on the first integrated design paradigm features to obtain the second integrated design paradigm features includes: Based on the first third space aggregation component among the plurality of third space aggregation components, the first first integrated design paradigm feature is processed by graph convolutional network encoding to obtain the first second integrated design paradigm feature; Based on the target third space aggregation component among the plurality of third space aggregation components, the target first integrated design paradigm feature is processed by graph convolutional network encoding to obtain the target second integrated design paradigm feature.

[0010] In one possible implementation, the method is executed based on a first architectural design generation model, which includes a spatial configuration generation module and a first feature enhancement module; the step of generating candidate architectural schemes based on the demand semantic features, using the design paradigm with the features of the second integrated design paradigm as generation constraints, includes: Based on the first feature enhancement module, feature enhancement processing is performed on the demand semantic features and the second integrated design paradigm features to obtain the features of the candidate building scheme. Based on the spatial configuration generation module, the spatial configuration decoding process is performed on the features of the candidate building scheme to obtain the candidate building scheme.

[0011] In one possible implementation, the first feature enhancement module includes a second visual dimensionality reduction mapping component and a graph convolution component, wherein the graph convolution component includes a fourth spatial aggregation component, a topological co-coding component, and a hierarchical coupling coding component; the feature enhancement processing based on the first feature enhancement module on the demand semantic features and the second integrated design paradigm features to obtain candidate architectural scheme features includes: Based on the second visual dimensionality reduction mapping component, the first design parameter perturbation feature is converted into the second visual dimensionality reduction feature, and the first design parameter perturbation feature is used to describe the morphological exploration factor; Based on the fourth spatial aggregation component, the second visual dimensionality reduction feature is processed by graph convolutional network encoding to obtain the third visual dimensionality reduction feature. Based on the topological co-coding component, structural co-aggregation modeling is performed on the third visual dimensionality reduction feature and the second integrated design paradigm feature to obtain the second transition feature; Based on the hierarchical coupled coding component, structural collaborative aggregation modeling is performed on the demand semantic features and the second transition features to obtain the features of the candidate building scheme. The first architectural design generation model further includes a first scale compression module, which includes multiple sequentially cascaded first feature compression components; the first feature enhancement module includes multiple sequentially cascaded first feature expansion components; and the second integrated design paradigm feature has multiple dimensions. The step of performing feature enhancement processing on the demand semantic features and the second integrated design paradigm features based on the first feature enhancement module to obtain the features of the candidate architectural scheme further includes: Based on the first first feature extension component among the plurality of first feature extension components, feature enhancement processing is performed on the first design parameter perturbation feature, the demand semantic feature, and the first second integrated design paradigm feature to obtain the first candidate building scheme feature. The first second integrated design paradigm feature is obtained based on the feature generated by the first first feature compression component. The first design parameter perturbation feature is used to describe the morphological exploration factor. Based on the target first feature extension component among the plurality of first feature extension components, feature enhancement processing is performed on the features of the preceding candidate building schemes, the semantic features of the requirements, and the features of the target second integrated design paradigm to obtain the features of the target candidate building schemes. The features of the target second integrated design paradigm are obtained based on the features generated by the target first feature compression component.

[0012] In one possible implementation, the method is performed based on a first architectural design generation model, the training process of which includes: Obtain architectural design requirement instances and benchmark architectural scheme instances, wherein the architectural design of the benchmark architectural scheme instances satisfies the design requirements defined by the architectural design requirement instances; Based on the second architectural design generation model, the sample requirement semantic features corresponding to the architectural design requirement instance are obtained, the first sample comparison features corresponding to the benchmark architectural scheme instance are obtained, and the second design parameter perturbation features are embedded in the first sample comparison features to obtain the second sample comparison features. The second design parameter perturbation features are used to describe the morphological exploration factor. Based on the second architectural design generation model, using the second sample comparison features as generation constraints, and based on the sample demand semantic features, generate sample candidate architectural schemes. Based on the sample candidate building schemes and the benchmark building scheme instance, the second building design generation model is trained to obtain the trained second building design generation model; Based on the trained second architectural design generation model, the first architectural design generation model is constructed. The first architectural design generation model that has completed training includes a sequentially cascaded feature extraction module, a second-scale compression module, a second-scale enhancement module, and a spatial configuration generation module. The second-scale compression module includes a sequentially cascaded first visual dimensionality reduction mapping component, a first spatial aggregation component, and a multi-scale fusion coding component. The second-scale enhancement module includes a sequentially cascaded second visual dimensionality reduction mapping component, a fourth spatial aggregation component, and a topological co-coding component. The construction of the first architectural design generation model based on the trained second architectural design generation model includes: By inheriting the first spatial aggregation component, a second spatial aggregation component is obtained; by inheriting the topology co-coding component, a hierarchical coupling coding component is obtained. By inheriting the fourth spatial aggregation component, a third spatial aggregation component is obtained; based on the feature integration component and the third spatial aggregation component, a feature integration module is constructed; In the second architectural design generation model that has completed training, the multi-scale fusion coding component of the second scale compression module is replaced with the second spatial aggregation component to obtain the first scale compression module. The hierarchical coupling coding component is embedded after the topological co-coding component of the second feature enhancement module to obtain the first feature enhancement module. The feature integration module is embedded between the first scale compression module and the first feature enhancement module to obtain the first architectural design generation model.

[0013] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being configured to perform any of the methods described in the first aspect.

[0014] Compared to existing technologies, the beneficial effects provided by this invention include: The method and system for intelligent generation of architectural design schemes based on deep learning, as disclosed in this invention, include: acquiring architectural design requirements and multiple comparative architectural schemes based on a consistent comparative design paradigm; extracting semantic features of the requirements and comparative design paradigm features of each comparative scheme; integrating the paradigm features of multiple comparative schemes and encoding them through a graph convolutional network to obtain a second integrated design paradigm feature; using this paradigm feature as a generation constraint, generating candidate architectural schemes that meet the design requirements based on the semantic features of the requirements. This invention improves the rationality and practicality of the generated schemes by integrating the consistent paradigm constraints of multiple comparative schemes. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the steps of the intelligent generation method for architectural design schemes based on deep learning provided in an embodiment of the present invention; Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0018] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the intelligent generation method for architectural design schemes based on deep learning provided in this embodiment. The following is a detailed description of this intelligent generation method for architectural design schemes based on deep learning.

[0020] Step S201: Obtain architectural design requirements and multiple comparative architectural schemes. The multiple comparative architectural schemes are based on a consistent comparative design paradigm. The architectural design requirements are used to describe the design requirements of the architectural design schemes. Step S202: Obtain the semantic features of the requirements corresponding to the architectural design requirements, and obtain the comparative design paradigm features corresponding to each of the multiple comparative architectural schemes. Step S203: Integrate the comparative design paradigm features corresponding to each of the multiple comparative architectural schemes to obtain the first integrated design paradigm feature; Step S204: Perform graph convolutional network encoding on the first integrated design paradigm feature to obtain the second integrated design paradigm feature, which is used to describe the comparative design paradigm. Step S205: Using the comparative design paradigm of the second integrated design paradigm features as the generation constraint, and based on the demand semantic features, generate candidate building schemes. The architectural design of the candidate building schemes meets the design requirements defined by the architectural design requirements.

[0021] In this embodiment of the invention, for example, this implementation uses a server as the execution entity to address the design requirements of a 120㎡ modern minimalist style three-bedroom apartment proposed by a real estate development company. The server first receives the structured requirements document uploaded by the company through the client. The document includes spatial function requirements (approximately 120㎡ three-bedroom apartment, master bedroom with ensuite bathroom and walk-in closet, open kitchen and dining room linkage, living room connected to a south-facing balcony, north-facing secondary bedroom with reserved space for conversion into a children's room and clear separation of active and quiet areas), style and indicator requirements (modern minimalist style facade with light gray paint + white horizontal lines, plot ratio ≤ 2.5, greening rate ≥ 30%, lighting coefficient ≥ 2.0%), and comparison scheme requirements (Lakeside Garden 118㎡ three-bedroom BIM model, City Light 122㎡ three-bedroom CAD drawings and renderings, Forest Mansion 120㎡ three-bedroom spatial topology model). The server retrieves the digital files of the above three comparative architectural schemes through the internal data interface and verifies their design paradigm consistency. After confirming that all three adopt the core layout logic of open kitchen + separation of active and quiet areas + south-facing balcony, they are used as valid input data.

[0022] Next, the server calls the pre-trained BERT-base semantic coding model to process the requirement document. First, the text is segmented, and then a 768-dimensional initial semantic vector is generated through the context coding layer. After being compressed to 512 dimensions through a fully connected layer, the requirement semantic feature F_sem is obtained. At the same time, the feature extraction module of the first architectural design generation model is started. The ResNet-50 convolutional neural network is used to extract 2048-dimensional architectural visual coding features from the digital files of the three comparison schemes. Then, the first visual dimensionality reduction mapping component of the first scale compression module converts it into 1024-dimensional first visual dimensionality reduction features. Subsequently, the first spatial aggregation component uses a graph convolutional network (GCN) to capture the spatial region connectivity to obtain the first transition feature. The second spatial aggregation component again uses GCN to strengthen the core paradigm elements, and finally obtains the 512-dimensional comparison design paradigm features F_par1-F_par3 for each comparison scheme.

[0023] The server continues to perform hierarchical integration of F_par1-F_par3 through the feature integration module. The first feature integration component uses an attention mechanism to weight and aggregate the basic spatial layout layer paradigm features (open kitchen weight 0.6, active and quiet zoning weight 0.4) to obtain the first first integration paradigm feature F_int1. The target feature integration component integrates the lighting and ventilation layer paradigm features to obtain the target first integration paradigm feature F_int2. The third feature integration component integrates the functional zoning layer paradigm features to obtain the third first integration paradigm feature F_int3. Then, multiple third spatial features in the feature integration module are called. The inter-space aggregation components perform graph convolutional encoding on these integrated features. The first third-space aggregation component captures the linear topological relationship between the open kitchen, dining room and living room through GCN to obtain the first second-space integration paradigm feature F_con1. The target third-space aggregation component strengthens the ventilation path between the south-facing balcony, living room and north-facing secondary bedroom to obtain the target second-space integration paradigm feature F_con2. The third third-space aggregation component optimizes the suite topology of the master bedroom, dressing room and bathroom to obtain the third second-space integration paradigm feature F_con3. Finally, the three are spliced ​​together to form a 1536-dimensional second integrated design paradigm feature F_con.

[0024] The server then initiates the first feature enhancement module, introducing a morphological exploration factor (allowing spatial dimensions to be adjusted ±5%). This is then converted into a 256-dimensional second visual dimensionality reduction feature F_dis by the second visual dimensionality reduction mapping component. The fourth spatial aggregation component uses GCN encoding to obtain the third visual dimensionality reduction feature F_dis3. The topological co-coding component uses a graph attention network (GAT) to structurally co-aggregate F_dis3 and F_con to obtain the second transition feature F_trans2. The hierarchical coupling encoding component again uses GAT to fuse the demand semantic feature F_sem with F_trans2 to obtain a 1024-dimensional candidate building scheme feature F_gen. Finally, the U-Net+3D convolutional layer of the spatial configuration generation module is called to decode F_gen, generating a 120.5m... 2 Floor plan (living room 25m) 2 The system connects a south-facing 3.6m wide balcony, an open L-shaped kitchen of 8㎡ with a 1.2m island, a master suite of 18㎡ + 5㎡ bathroom + 4㎡ walk-in closet, and a north-facing secondary bedroom with a reserved renovation interface. It also generates modern minimalist style facade drawings (light gray real stone paint with white horizontal lines) and interior and exterior 3D renderings. At the same time, it automatically calculates and generates a plan with a plot ratio of 2.4, a greening rate of 32%, and a lighting coefficient of 2.1%. After confirming that the requirements and constraints are met, the candidate plan is packaged and returned to the enterprise client. The enterprise can make modification suggestions, and the server will iterate and generate a new plan again.

[0025] This embodiment fully covers all the technical steps described in this invention. Through specific architectural design requirements, comparative schemes, and generated results, it clearly demonstrates how the server uses deep learning technology to generate intelligent building schemes under the constraints of requirements and paradigms, reflecting the practicality and technological advancement of the method.

[0026] In this embodiment of the invention, the method is executed based on a first architectural design generation model, which includes a feature extraction module and a first scale compression module. The step of obtaining the comparative design paradigm features corresponding to each of the multiple comparative architectural schemes can be implemented through the following example.

[0027] Based on the feature extraction module, feature extraction processing is performed on each of the comparative architectural schemes to obtain the architectural visual coding features corresponding to each of the comparative architectural schemes. Based on the first scale compression module, feature compression is performed on the architectural visual coding features corresponding to each of the comparative architectural schemes to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes.

[0028] In an embodiment of the invention, for example, the server initiates the feature extraction module of the first architectural design generation model for three comparison schemes: a 118㎡ BIM model of Lakeside Garden, a 122㎡ CAD rendering of City of Light, and a 120㎡ topology model of Forest Mansion. The 3D spatial data of the Lakeside Garden BIM model is converted into a 2D layout diagram and input into a ResNet-50 convolutional network to extract 2048-dimensional architectural visual coding features F_vis1, such as the outline of the open kitchen island and the boundary of the dynamic and static partition walls. The CAD rendering of City of Light is pixelated and input into ResNet-50 to extract features such as the light gray facade and white horizontal lines to obtain F_vis2. The topology model of Forest Mansion is converted into a functional area connection diagram and input into ResNet-50 to extract the connection features of the master bedroom-walk-in closet-bathroom to obtain F_vis3. Next, the first-scale compression module is called: F_vis1-3 is compressed to 1024-dimensional first visual dimensionality reduction features F_red1-3 through the first visual dimensionality reduction mapping component (fully connected layer); the first spatial aggregation component constructs a graph structure with functional areas as nodes and connectivity as edges, and performs GCN encoding on F_red1 to obtain the first transition feature F_trans1 that captures the spatial topology; the second spatial aggregation component further strengthens the core paradigm of the L-shaped kitchen + island through GCN, and outputs a 512-dimensional comparative design paradigm feature F_par1; similarly, F_red2 is processed to obtain F_par2 which strengthens the linkage of the two balconies, and F_red3 is processed to obtain F_par3 which strengthens the separation of the master bedroom suite.

[0029] In this embodiment of the invention, the first-scale compression module includes a first visual dimensionality reduction mapping component and a graph convolution processing component; the feature compression of the architectural visual coding features corresponding to each of the comparative architectural schemes based on the first-scale compression module to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes can be implemented through the following example.

[0030] Based on the first visual dimensionality reduction mapping component, the architectural visual coding features corresponding to each of the comparative architectural schemes are converted into the first visual dimensionality reduction features corresponding to each of the comparative architectural schemes. Based on the graph convolution processing component, spatial aggregation encoding processing is performed on the first visual dimensionality reduction features corresponding to each of the comparative architectural schemes to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes.

[0031] In an embodiment of the invention, for example, the server activates the first visual dimensionality reduction mapping component of the first scale compression module for the 2048-dimensional architectural visual coding features F_vis1, F_vis2, and F_vis3 corresponding to the 118㎡ BIM model of Lakeside Garden, the 122㎡ CAD rendering of City of Light, and the 120㎡ topological model of Forest Mansion. This component is a pre-trained fully connected layer that compresses F_vis1, which contains elements such as an L-shaped kitchen island and a wall separating active and quiet zones, to a 1024-dimensional first visual dimensionality reduction feature F_red1; compresses F_vis2, which contains elements such as a light gray facade and white horizontal lines, to a 1024-dimensional F_red2; and compresses F_vis3, which contains the connection relationship between the master bedroom, walk-in closet, and bathroom, to a 1024-dimensional F_red3, while retaining the core visual and topological information.

[0032] Then, the graph convolution processing component is invoked to perform spatial aggregation encoding: For F_red1, the component first constructs a functional area graph structure (nodes are living room, kitchen, master bedroom, etc., and edges are the connectivity relationships between functional areas), captures the topological relationships of functional areas through the first layer of GCN to obtain transition features, and then strengthens the core paradigm of "L-shaped kitchen + island" through the second layer of GCN, outputting a 512-dimensional comparative design paradigm feature F_par1; For F_red2, the component constructs an exterior facade element graph (nodes are walls, lines, balconies, and edges are positional relationships), and after aggregation, strengthens the "double balcony linkage" feature to obtain F_par2; For F_red3, the component constructs a suite functional area graph (nodes are master bedroom, walk-in closet, bathroom, and edges are internal connectivity relationships), and after aggregation, strengthens the "master bedroom suite separation" feature to obtain F_par3. All steps are automatically executed by the server, outputting feature vectors that conform to the comparative design paradigm.

[0033] In this embodiment of the invention, the graph convolution processing component includes a first spatial aggregation component and a second spatial aggregation component; the step of performing spatial aggregation encoding processing on the first visual dimensionality reduction features corresponding to each of the comparative architectural schemes based on the graph convolution processing component to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes can be implemented through the following example.

[0034] Based on the first spatial aggregation component, graph convolutional network encoding is performed on the first visual dimensionality reduction features corresponding to each of the comparative architectural schemes to obtain the first transition features corresponding to each of the comparative architectural schemes. Based on the second spatial aggregation component, graph convolutional network encoding is performed on the first transition features corresponding to each of the comparative architectural schemes to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes.

[0035] In an embodiment of the invention, for example, the server uses the 1024-dimensional first visual dimensionality reduction features F_red1, F_red2, and F_red3 corresponding to the 118㎡ BIM model of Lakeside Garden, the 122㎡ CAD rendering of City of Light, and the 120㎡ topological model of Forest Mansion to initiate the first spatial aggregation component of the graph convolution processing component: For F_red1 (including L-shaped kitchen island and dynamic / static zoning elements), the component constructs a functional area graph structure, with nodes such as living room, L-shaped kitchen, island, master bedroom, etc., and edges representing the adjacent relationships of functional areas (such as kitchen connected to island, island connected to dining room). Through the GCN encoding layer, the features of each node are aggregated with the features of adjacent nodes to output the first transition feature F_trans1 that strengthens the topological association of functional areas, clearly capturing the coherent layout logic of kitchen-island-dining room.

[0036] For F_red2 (which includes two balconies and light gray facade elements), the component constructs a facade + space diagram. The nodes are the south-facing balcony, the north-facing balcony, the living room, and the secondary bedroom, and the edges are the lighting and ventilation paths (such as the south-facing balcony connecting to the living room and the north-facing balcony connecting to the secondary bedroom). After GCN encoding, F_trans2 is output to highlight the lighting linkage feature of the two balconies.

[0037] For F_red3 (containing the master bedroom suite element), the component constructs the master bedroom suite graph, with nodes representing the master bedroom, walk-in closet, and ensuite bathroom, and edges representing internal passage relationships; after GCN encoding, it outputs F_trans3, which reflects the independent partitioning structure of the suite.

[0038] Then the second space aggregation component is activated: For F_trans1, the component focuses on the core paradigm "L-shaped kitchen + island" through the attention-weighted GCN layer, giving higher weights to the kitchen and island nodes, and outputs 512-dimensional comparative design paradigm features F_par1 after aggregation; For F_trans2, focus on the "double balcony linkage" paradigm, weight the two balcony nodes, and output F_par2; For F_trans3, focusing on the "master suite separation" paradigm, weighting is applied to the walk-in closet and bathroom nodes, outputting F_par3. All steps are completed automatically by the server, and the feature vectors accurately retain the core design paradigm of each comparison scheme.

[0039] In this embodiment of the invention, the first scale compression module includes a plurality of first feature compression components cascaded in sequence; the feature compression of the architectural visual coding features corresponding to each of the comparative architectural schemes based on the first scale compression module to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes can be implemented through the following example.

[0040] Based on the first first feature compression component among the plurality of first feature compression components, feature compression is performed on the architectural visual coding features corresponding to each of the comparative architectural schemes to obtain the first comparative design paradigm feature corresponding to each of the comparative architectural schemes. Based on the target first feature compression component among the plurality of first feature compression components, feature compression is performed on the preceding comparative design paradigm features corresponding to each of the comparative architectural schemes to obtain the target comparative design paradigm features corresponding to each of the comparative architectural schemes, wherein the preceding comparative design paradigm feature is the one preceding the target first feature compression component.

[0041] In an embodiment of the invention, for example, the server initiates a first-scale compression module, which includes three sequentially cascaded first feature compression components (component 1, component 2, and component 3), performing cascaded compression on the 2048-dimensional architectural visual encoding features of Lakeside Garden (F_vis1), City Light (F_vis2), and Forest Mansion (F_vis3): First, the first component (Component 1) processes the visual encoding features of each comparison scheme: For F_vis1 (Lakeside Garden), Component 1 first reduces the dimensionality from 2048 to 1024 through a fully connected layer, and then captures the basic spatial layout relationship of the living room-kitchen-dining room through a basic graph convolutional network (GCN), outputting the first comparative design paradigm feature F_par1_1 (512 dimensions); For F_vis2 (City Light), after Component 1 reduces the dimensionality to 1024, GCN captures the lighting linkage between the double balconies and the living room, outputting F_par2_1 (512 dimensions); For F_vis3 (Forest Mansion), after Component 1 reduces the dimensionality to 1024, GCN captures the suite structure of the master bedroom-walk-in closet-bathroom, outputting F_par3_1 (512 dimensions).

[0042] Next, the target component (component 2) processes the preceding paradigm features: For F_par1_1, component 2 focuses on the core paradigm of "L-shaped kitchen + island" through attention-weighted GCN, compressing 512 dimensions to 256 dimensions, and outputs the target paradigm feature F_par1_2 (256 dimensions); For F_par2_1, component 2 focuses on "double balconies with north-south linkage", compressing it to 256 dimensions to output F_par2_2; For F_par3_1, component 2 focuses on "master bedroom suite with independent zoning", compressing it to 256 dimensions to output F_par3_2.

[0043] Finally, another target component (component 3) processes the preceding features F_par1_2 / F_par2_2 / F_par3_2: For F_par1_2, component 3 further compresses it to 128 dimensions using a multilayer perceptron, and then strengthens the topological coupling between the island and the kitchen through a refined GCN, outputting the final comparative design paradigm feature F_par1 (128 dimensions); for F_par2_2, component 3 strengthens the ventilation path coupling of the two balconies, outputting F_par2 (128 dimensions); for F_par3_2, component 3 strengthens the circulation connection between the walk-in closet and the bathroom, outputting F_par3 (128 dimensions).

[0044] All steps are executed automatically by the server. Cascaded components progressively compress dimensions and strengthen core paradigm features, ultimately outputting accurate paradigm vectors for each comparison scheme.

[0045] In this embodiment of the invention, the first architectural design generation model includes a feature integration module, which includes multiple feature integration components; the integration of the comparative design paradigm features corresponding to the multiple comparative architectural schemes to obtain the first integrated design paradigm feature can be implemented through the following example.

[0046] Based on the first feature integration component among the multiple feature integration components, the first comparative design paradigm features corresponding to each of the multiple comparative architectural schemes are integrated to obtain the first integrated design paradigm feature; Based on the target feature integration component among the multiple feature integration components, the target comparative design paradigm features corresponding to each of the multiple comparative architectural schemes are integrated to obtain the target first integrated design paradigm features.

[0047] In an embodiment of the present invention, for example, the server initiates a feature integration module for a first architectural design generation model, which includes a first feature integration component and a target feature integration component.

[0048] First, the initial feature integration component receives features from three initial comparative design paradigms: Lakeside Garden's F_par1_1 (512 dimensions, including an L-shaped kitchen and island), City Light's F_par2_1 (512 dimensions, including a linked double balcony), and Forest Mansion's F_par3_1 (512 dimensions, including a separate master suite). The component uses an attention mechanism to align feature dimensions, assigning weights based on the commonalities of modern minimalist style: F_par1_1 accounts for 30%, F_par2_1 for 35%, and F_par3_1 for 35%. After weighted summation, the first integrated design paradigm feature F_int1 (512 dimensions) is output, merging the core basic paradigms of the three.

[0049] Next, the target feature integration component receives three comparative design paradigm features from three targets: Lakeside Garden (F_par1_2, 256-dimensional, emphasizing the L-shaped kitchen and island), City Light (F_par2_2, 256-dimensional, emphasizing the linkage between the two balconies), and Forest Mansion (F_par3_2, 256-dimensional, emphasizing the separation of the master suite). The component constructs these three as graph nodes, with edges representing spatial relationships between paradigms (such as the coupling of circulation between the kitchen and balcony). It aggregates the node features through a graph convolutional network and outputs the first integrated design paradigm feature F_int2 (256-dimensional), strengthening the paradigm synergy. All operations are automatically executed by the server, with consistent feature dimensions to support subsequent processing.

[0050] In this embodiment of the invention, the feature integration module further includes multiple third-space aggregation components; the step of performing graph convolutional network encoding on the first integrated design paradigm features to obtain the second integrated design paradigm features can be implemented through the following example.

[0051] Based on the first third space aggregation component among the plurality of third space aggregation components, the first first integrated design paradigm feature is processed by graph convolutional network encoding to obtain the first second integrated design paradigm feature; Based on the target third space aggregation component among the plurality of third space aggregation components, the target first integrated design paradigm feature is processed by graph convolutional network encoding to obtain the target second integrated design paradigm feature.

[0052] In an embodiment of the invention, for example, the server initiates the third space aggregation component in the feature integration module to process the first generated first integrated design paradigm feature F_int1 (512 dimensions, integrating three basic paradigms: L-shaped kitchen + island, double balcony linkage, and master bedroom suite separation) and the target first integrated design paradigm feature F_int2 (256 dimensions, strengthening paradigm synergy relationships) respectively: First, the initial third-space aggregation component receives F_int1 and decomposes it into three core paradigm nodes. Node attributes include the spatial parameters of each paradigm (e.g., the aspect ratio of an L-shaped kitchen is 1.5:1, and the north-south distance between the two balconies is 3.2 meters), and edges represent the spatial dependencies between nodes (e.g., the complementary lighting between the kitchen island and the south-facing balcony, and the alignment of the master suite with the boundaries of the active and quiet zones). The component uses a graph convolutional network (GCN) to weight and aggregate the node features, merging the attributes of adjacent nodes into a unified spatial logic, and outputting the first second integrated design paradigm feature F_con1 (512-dimensional, enhancing the spatial synergy of the three paradigms).

[0053] Next, the target third-space aggregation component receives F_int2 and maps it to more refined sub-paradigm nodes (such as the extension direction of the island, the window area of ​​the balcony (1.8㎡), and the circulation length of the walk-in closet (2.5 meters). The edges represent the parameter constraints between the nodes (such as the island extension direction needing to match the linkage between the kitchen and dining room, and the window area needing to meet a daylighting coefficient ≥ 2.0%). The component uses the attention mechanism GCN, assigning a high weight of 0.4 to the balcony window nodes that affect daylighting. After aggregation, it outputs the target second integrated design paradigm feature F_con2 (256 dimensions, precisely encoding the parameter constraint boundaries of each sub-paradigm). All operations are automatically executed by the server, and the output features can be directly used as the constraint basis for subsequent scheme generation.

[0054] In this embodiment of the invention, the method is executed based on a first architectural design generation model, which includes a spatial configuration generation module and a first feature enhancement module. The method of generating candidate architectural schemes based on the second integrated design paradigm features as generation constraints and the demand semantic features can be implemented through the following example.

[0055] Based on the first feature enhancement module, feature enhancement processing is performed on the demand semantic features and the second integrated design paradigm features to obtain the features of the candidate building scheme. Based on the spatial configuration generation module, the spatial configuration decoding process is performed on the features of the candidate building scheme to obtain the candidate building scheme.

[0056] In an embodiment of the invention, for example, the server initiates the first feature enhancement module and the spatial configuration generation module of the first architectural design generation model to execute the generation process for a 120㎡ modern minimalist three-bedroom apartment: First, the first feature enhancement module receives the semantic feature F_sem (512 dimensions, encoding core requirements such as a 120㎡ three-bedroom apartment, modern minimalist style, south-facing balcony, and separation of active and quiet zones) and the second integrated design paradigm feature F_con (768 dimensions, composed of F_con1 (512 dimensions, integrating paradigms such as L-shaped kitchen + island, and double balcony linkage) and F_con2 (256 dimensions, encoding constraints such as daylighting coefficient ≥2.0%)). The module fuses the two through an attention mechanism to enhance the matching between requirements and paradigms (such as adapting to the L-shaped kitchen size of a 120㎡ space) and outputs the candidate architectural scheme feature F_gen (1024 dimensions).

[0057] Next, the spatial configuration generation module receives F_gen and uses U-Net to decode and generate a 2D floor plan: 120.5㎡, a 25㎡ living room connected to a 3.6m wide south-facing balcony, an 8㎡ L-shaped kitchen with a 1.2m island, a master suite (18㎡ bedroom + 5㎡ bathroom + 4㎡ walk-in closet), and a 12㎡ north-facing secondary bedroom with a reserved interface for a children's room; the exterior facade (light gray stone paint + white lines with 0.8m spacing) and interior renderings are generated through 3D convolution; the plot ratio of 2.4, the greening rate of 32%, and the daylight factor of 2.1% are automatically verified to meet the standards, and finally, a complete candidate building scheme is output.

[0058] In this embodiment of the invention, the first feature enhancement module includes a second visual dimensionality reduction mapping component and a graph convolution component; the feature enhancement processing of the demand semantic features and the second integrated design paradigm features based on the first feature enhancement module to obtain the features of the candidate building scheme can be implemented through the following example.

[0059] Based on the second visual dimensionality reduction mapping component, the first design parameter perturbation feature is converted into the second visual dimensionality reduction feature, and the first design parameter perturbation feature is used to describe the morphological exploration factor; Based on the graph convolutional component, spatial aggregation encoding is performed on the demand semantic features, the second integrated design paradigm features, and the second visual dimensionality reduction features to obtain the features of the candidate building scheme.

[0060] In an embodiment of the invention, for example, the server, targeting the needs of a 120㎡ modern minimalist three-bedroom apartment, activates the second visual dimensionality reduction mapping component and the graph convolution component of the first feature enhancement module: First, the server generates a first design parameter perturbation feature (morphological exploration factor): a 64-dimensional parameter vector containing a 0.15m increase in island length, a 0.2m decrease in south-facing balcony width, and a 0.3m extension in the master bedroom walk-in closet circulation path. This feature is used to explore morphological variants that meet the constraints. The second visual dimensionality reduction mapping component converts it into a 32-dimensional second visual dimensionality reduction feature F_dis2 through a pre-trained fully connected layer, preserving the core perturbation direction and adapting it to other feature dimensions.

[0061] Next, the graph convolutional component receives the semantic features of the requirements F_sem (512-dimensional, encoding core requirements such as 120㎡ and modern minimalist style), the second integrated design paradigm features F_con (768-dimensional, integrating L-shaped kitchen + island, double balcony linkage and lighting constraints), and the second visual dimensionality reduction features F_dis2 (32-dimensional). The component constructs a graph structure: nodes represent the core requirements (120㎡ space, minimalist style), the core paradigm (L-shaped kitchen proportions, double balcony lighting parameters), and the core perturbations (island adjustment, balcony adjustment); edges represent the constraint relationships between nodes (e.g., 120㎡ island adjustment is limited to no more than 0.2m, double balcony lighting constraint reduces balcony width by no less than 3.4m). An attention graph convolutional network (GAT) is used, assigning a high weight of 0.35 to perturbation nodes that conform to the paradigm constraints, aggregating the three features into a unified spatial logic, and outputting the candidate building scheme feature F_gen (1024-dimensional). This feature satisfies both the requirements and paradigm constraints, and incorporates reasonable form exploration.

[0062] In this embodiment of the invention, the graph convolution component includes a fourth spatial aggregation component, a topological co-coding component, and a hierarchical coupling coding component; the spatial aggregation coding process performed on the demand semantic features, the second integrated design paradigm features, and the second visual dimensionality reduction features based on the graph convolution component to obtain the candidate building scheme features can be implemented through the following example.

[0063] Based on the fourth spatial aggregation component, the second visual dimensionality reduction feature is processed by graph convolutional network encoding to obtain the third visual dimensionality reduction feature. Based on the topological co-coding component, structural co-aggregation modeling is performed on the third visual dimensionality reduction feature and the second integrated design paradigm feature to obtain the second transition feature; Based on the hierarchical coupling coding component, structural collaborative aggregation modeling is performed on the demand semantic features and the second transition features to obtain the features of the candidate building scheme.

[0064] In an embodiment of the invention, for example, the server, targeting the needs of a 120㎡ modern minimalist three-bedroom apartment, initiates three sub-components of the graph convolutional component: fourth spatial aggregation, topological co-coding, and hierarchical coupling coding. First, the fourth spatial aggregation component receives the second visual dimensionality reduction feature F_dis2 (32-dimensional, including morphological perturbation parameters of island platform +0.15m, south-facing balcony -0.2m, and cloakroom circulation +0.3m). The component maps F_dis2 into three perturbation nodes (island platform adjustment, balcony adjustment, and cloakroom adjustment), with edges representing parameter constraints (e.g., connecting the balcony adjustment node to the virtual node with "lighting coefficient ≥2.0%"). The component strengthens the rationality of the perturbation through GCN encoding (verifying that the balcony is still ≥3.4m after -0.2m), and outputs the third visual dimensionality reduction feature F_dis3 (64-dimensional).

[0065] Next, the topology co-coding component receives F_dis3 and the second integrated design paradigm feature F_con (768-dimensional, integrating paradigms such as the 1.5:1 ratio of the L-shaped kitchen and the linkage of double balconies). The component constructs a co-coding graph: the nodes include 3 perturbation nodes of F_dis3 and 4 paradigm core nodes of F_con (L-shaped kitchen ratio, double balcony lighting, master bedroom suite circulation, and dynamic and static zoning boundaries), and the edges are adaptation relationships (e.g., the island adjustment node is connected to the L-shaped kitchen ratio node, and it is verified that the 1.5:1 ratio is still met after the length is increased by 0.15m). Using the GAT attention mechanism, the adaptation edges are assigned a high weight of 0.4, and the aggregation is used to obtain the second transition feature F_trans2 (1024-dimensional), realizing the co-coding of perturbations and paradigms.

[0066] Finally, the hierarchical coupled coding component receives the semantic features of requirements F_sem (512-dimensional, encoding core requirements such as 120㎡ space, modern minimalist style, and south-facing balcony) and F_trans2. The component constructs a hierarchical graph: the upper layer consists of 2 requirement nodes (120㎡ space constraint and minimalist style constraint), and the lower layer consists of 7 collaborative nodes of F_trans2; the edges represent matching relationships (e.g., connecting the 120㎡ node with the island adjustment node, verifying that the total kitchen area after lengthening is ≤8.5㎡). The hierarchical GCN first aggregates the lower-level collaborative nodes and then merges them with the upper-level requirement nodes, outputting the candidate building scheme feature F_gen (1024-dimensional). This feature satisfies both requirement and paradigm constraints while incorporating reasonable form exploration.

[0067] In this embodiment of the invention, the first architectural design generation model further includes a first scale compression module, which includes a plurality of sequentially cascaded first feature compression components, and a first feature enhancement module, which includes a plurality of sequentially cascaded first feature extension components. The second integrated design paradigm feature has multiple dimensions. The feature enhancement processing performed on the demand semantic feature and the second integrated design paradigm feature based on the first feature enhancement module to obtain the feature of the candidate architectural scheme can be implemented through the following example.

[0068] Based on the first first feature extension component among the plurality of first feature extension components, feature enhancement processing is performed on the first design parameter perturbation feature, the demand semantic feature, and the first second integrated design paradigm feature to obtain the first candidate building scheme feature. The first second integrated design paradigm feature is obtained based on the feature generated by the first first feature compression component. The first design parameter perturbation feature is used to describe the morphological exploration factor. Based on the target first feature extension component among the plurality of first feature extension components, feature enhancement processing is performed on the features of the preceding candidate building schemes, the semantic features of the requirements, and the features of the target second integrated design paradigm to obtain the features of the target candidate building schemes. The features of the target second integrated design paradigm are obtained based on the features generated by the target first feature compression component.

[0069] In an embodiment of the invention, for example, the server, targeting a 120㎡ modern minimalist three-bedroom apartment, initiates two sequentially cascaded expansion components (first expansion component E1, target expansion component E2) of the first feature enhancement module, and performs feature enhancement in conjunction with the multi-dimensional second integrated design paradigm features (F_con1, F_con2) generated by the first scale compression module: First, the server generates the first design parameter perturbation feature (morphological exploration factor): a 64-dimensional vector that encodes morphological variations of the island length +0.15m, the south-facing balcony width -0.2m, and the master bedroom walk-in closet circulation +0.3m, which are used to explore reasonable adjustment space within the paradigm constraints.

[0070] Next, the first extended component E1 receives the perturbation feature, the demand semantic feature F_sem (512-dimensional, encoding core needs such as 120㎡ space, modern minimalist style, and south-facing balcony), and the first second integrated paradigm feature F_con1 (512-dimensional, from the integrated encoding of the comparison scheme features generated by the first scale compression component 1, including basic paradigms such as L-shaped kitchen with a 1.5:1 ratio and north-south distance of 3.2m between double balconies). Within E1, the fourth spatial aggregation component maps the disturbance features to three nodes (island adjustment, balcony adjustment, and cloakroom adjustment), with edges constrained by parameters (e.g., balcony adjustment ≤ 0.2m to meet daylighting coefficient ≥ 2.0%). The topology co-coding component constructs adaptation edges between the nodes and the four paradigm core nodes of F_con1 (L-shaped kitchen ratio, double balcony linkage, etc.) (e.g., the island adjustment is connected to the L-shaped kitchen ratio, and the lengthening still meets the 1.5:1 ratio). GAT is used to assign a weight of 0.4 to the adaptation edges. The hierarchical coupling coding component merges the aggregation results with F_sem and outputs the first candidate building scheme feature F_gen1 (1024 dimensions), realizing the synergy between disturbance and basic paradigm.

[0071] Subsequently, the target extension component E2 receives F_gen1, F_sem, and the target second integration paradigm feature F_con2 (256 dimensions, from the feature integration encoding of the comparison scheme generated by the first scale compression component 2, including sub-paradigm parameter constraints: balcony window area ≥ 1.8㎡, walk-in closet circulation ≤ 2.8m). E2 internally constructs a collaboration graph: nodes include 7 collaboration nodes of F_gen1 and 3 sub-paradigm constraint nodes of F_con2 (balcony window opening, walk-in closet circulation, and dynamic / static zoning boundary); edges represent parameter verification relationships (e.g., connecting the balcony adjustment node with the window area node verifies that the window opening is still ≥ 1.8㎡ after -0.2m); a hierarchical GCN is used to first aggregate the lower-level collaboration nodes, then merge them with the upper-level F_sem requirement nodes, outputting the target candidate building scheme feature F_gen2 (1024 dimensions). This feature satisfies both requirements and multi-dimensional paradigm constraints, while also incorporating reasonable form exploration, providing a precise basis for subsequent spatial configuration generation.

[0072] In this embodiment of the invention, the spatial configuration decoding process performed on the features of the candidate building scheme based on the spatial configuration generation module to obtain the candidate building scheme can be implemented through the following example.

[0073] Based on the spatial configuration generation module, the spatial configuration decoding process is performed on the features of the final candidate building scheme to obtain the candidate building scheme.

[0074] In this embodiment of the invention, for example, the server receives the final candidate building scheme feature F_gen2 (1024-dimensional, integrating morphological exploration, demand semantics, and multi-dimensional paradigm constraints) and initiates the spatial configuration generation module. The module decodes using U-Net: downsampling to extract the core logic of 120.5㎡ functional zoning, upsampling to generate a 2D floor plan, including a 25㎡ living room connected to a 3.4m wide south-facing balcony (adjusted), an 8.2㎡ L-shaped kitchen with a 1.35m extended island, a master suite including an 18㎡ bedroom, a 5㎡ bathroom, and a 4.3㎡ walk-in closet with extended circulation, and a 12㎡ north-facing secondary bedroom with a reserved interface for a children's room. Then, a 3D convolutional layer expands the 2D features into a three-dimensional space, rendering a modern minimalist style facade (light gray stone paint + 0.8m white horizontal lines) and interior renderings (with minimalist furniture layout). Finally, it automatically verifies that the plot ratio of 2.4, the greening rate of 32%, and the daylight factor of 2.1% all meet the requirements, and outputs a complete candidate building scheme including floor plans, renderings, and an indicator report.

[0075] In this embodiment of the invention, the method is executed based on a first architectural design generation model, and the training process of the first architectural design generation model can be implemented through the following example.

[0076] Obtain architectural design requirement instances and benchmark architectural scheme instances, wherein the architectural design of the benchmark architectural scheme instances satisfies the design requirements defined by the architectural design requirement instances; Based on the second architectural design generation model, the sample requirement semantic features corresponding to the architectural design requirement instance are obtained, the first sample comparison features corresponding to the benchmark architectural scheme instance are obtained, and the second design parameter perturbation features are embedded in the first sample comparison features to obtain the second sample comparison features. The second design parameter perturbation features are used to describe the morphological exploration factor. Based on the second architectural design generation model, using the second sample comparison features as generation constraints, and based on the sample demand semantic features, generate sample candidate architectural schemes. Based on the sample candidate building schemes and the benchmark building scheme instance, the second building design generation model is trained to obtain the trained second building design generation model; The first architectural design generation model is constructed based on the second architectural design generation model that has been trained.

[0077] In an embodiment of the present invention, for example, the server initiates the training process of the first architectural design generation model. First, it retrieves an example of architectural design requirements (120㎡ modern minimalist three-bedroom apartment: master bedroom with ensuite bathroom and dressing room, open kitchen and connected dining room, south-facing balcony, separation of active and quiet areas, plot ratio ≤2.5, daylighting coefficient ≥2.0%) and the corresponding benchmark scheme example (Lakeside Garden 118㎡ three-bedroom apartment BIM model has been built, which meets the requirements).

[0078] Next, the server calls the second architectural design generation model, and uses its feature extraction module to generate 512-dimensional sample semantic features of the requirement instance; it extracts 2048-dimensional first sample comparison features (including L-shaped kitchen 1.5:1 ratio, north-south distance of double balconies 3.2m, etc.) from the benchmark scheme instance, embeds the second design parameter disturbance features (64-dimensional, including morphological exploration factors of island length ±0.2m and south-facing balcony width ±0.3m), and fuses them to obtain 2112-dimensional second sample comparison features.

[0079] Then, the second model, using this feature as a constraint and the semantic features of the sample requirements as a basis, generates sample candidate solutions (a floor plan and rendering after adjusting the island length to 1.35m). The server calculates the loss between the sample candidate solutions and the benchmark solution: spatial layout error (kitchen area deviation ≤ 0.5㎡), style matching degree (modern minimalist element coverage ≥ 90%), and indicator compliance rate (floor area ratio of 2.4 meets the requirements). The parameters of the second model are optimized through backpropagation until the loss is lower than the preset threshold (0.05), thus obtaining the trained second model.

[0080] Finally, based on this model, the server inherits its core structure, such as the feature extraction module and the fourth spatial aggregation component, and adds a multi-scale feature integration module (used for hierarchical integration of features of comparative schemes) to construct the first architectural design generation model.

[0081] In this embodiment of the invention, the first architectural design generation model that has completed training includes a sequentially cascaded feature extraction module, a second scale compression module, a second feature enhancement module, and a spatial configuration generation module. The second scale compression module includes a sequentially cascaded first visual dimensionality reduction mapping component, a first spatial aggregation component, and a multi-scale fusion coding component. The second feature enhancement module includes a sequentially cascaded second visual dimensionality reduction mapping component, a fourth spatial aggregation component, and a topological co-coding component. The construction of the first architectural design generation model based on the trained second architectural design generation model can be implemented through the following example.

[0082] By inheriting the first spatial aggregation component, a second spatial aggregation component is obtained; by inheriting the topology co-coding component, a hierarchical coupling coding component is obtained. By inheriting the fourth spatial aggregation component, a third spatial aggregation component is obtained; based on the feature integration component and the third spatial aggregation component, a feature integration module is constructed; In the second architectural design generation model that has completed training, the multi-scale fusion coding component of the second scale compression module is replaced with the second spatial aggregation component to obtain the first scale compression module. The hierarchical coupling coding component is embedded after the topological co-coding component of the second feature enhancement module to obtain the first feature enhancement module. The feature integration module is embedded between the first scale compression module and the first feature enhancement module to obtain the first architectural design generation model.

[0083] In an embodiment of the invention, for example, the server loads the trained second architectural design generation model (trained for a 120㎡ modern minimalist three-bedroom apartment scene, the structure of which includes a feature extraction module, a second scale compression module, a second feature enhancement module, and a spatial configuration generation module), and initiates the construction process of the first architectural design generation model: First, the server inherits the core components: it copies the first spatial aggregation component (GCN structure, trained to capture the spatial topology of the comparison scheme) of the second scale compression module in the second model, retains its topological encoding capability of "functional area as node and connectivity as edge", and names it the second spatial aggregation component; it copies the topological co-coding component (GAT structure, trained with paradigm and perturbation adaptation weights) of the second feature enhancement module, adjusts the input dimension to adapt to the semantic feature fusion requirements, and names it the hierarchical coupling coding component.

[0084] Next, the server constructs a feature integration module: it copies the fourth spatial aggregation component (GCN structure) of the second feature enhancement module and names it the third spatial aggregation component; it creates a feature integration component (attention weighting layer, with preset weight rules of "open kitchen accounts for 30%, double balcony linkage accounts for 35%, and master bedroom suite separation accounts for 35%), and connects the feature integration component with the third spatial aggregation component. First, it integrates the paradigm features of multiple comparison schemes through attention, and then the spatial association of the integrated features is encoded by the third spatial aggregation component to form the feature integration module.

[0085] Then, the server modifies the second model structure: the multi-scale fusion coding component in the second-scale compression module is replaced with the second spatial aggregation component to obtain the first-scale compression module (in the order: first visual dimensionality reduction mapping → first spatial aggregation → second spatial aggregation), which enhances the topological refinement of the contrast scheme paradigm features; a hierarchical coupling coding component is embedded after the topological co-coding component of the second feature enhancement module to obtain the first feature enhancement module (in the order: second visual dimensionality reduction mapping → fourth spatial aggregation → topological co-coding → hierarchical coupling coding), which realizes the hierarchical fusion of demand semantics and transition features; finally, a feature integration module is inserted between the first-scale compression module and the first feature enhancement module to form a complete first architectural design generation model (in the order: feature extraction → first-scale compression → feature integration → first feature enhancement → spatial configuration generation).

[0086] The server saves the structure and parameters of the first model, completes the construction, and can be directly used for the intelligent generation of a 120㎡ modern minimalist three-bedroom apartment design.

[0087] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent generation method for architectural design schemes based on deep learning. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other based on one or more communication buses or signal lines.

[0088] For illustrative purposes, the foregoing description has been made in comparison to specific embodiments. However, the foregoing illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for intelligent generation of architectural design schemes based on deep learning, characterized in that, include: Obtain architectural design requirements and multiple comparative architectural schemes, the multiple comparative architectural schemes being based on a consistent comparative design paradigm, and the architectural design requirements being used to describe the design requirements of the architectural design schemes. Obtain the semantic features of the architectural design requirements, and obtain the comparative design paradigm features of each comparative architectural scheme in the multiple comparative architectural schemes. By integrating the comparative design paradigm features corresponding to the multiple comparative architectural schemes, the first integrated design paradigm feature is obtained; The first integrated design paradigm feature is processed by graph convolutional network encoding to obtain the second integrated design paradigm feature, which is used to describe the comparative design paradigm. Using the comparative design paradigm of the second integrated design paradigm features as the generation constraint, and based on the requirement semantic features, candidate architectural schemes are generated, and the architectural design of the candidate architectural schemes meets the design requirements defined by the architectural design requirements. The method is executed based on a first architectural design generation model, which includes a feature extraction module and a first scale compression module. The step of obtaining the comparative design paradigm features corresponding to each of the multiple comparative architectural schemes includes: Based on the feature extraction module, feature extraction processing is performed on each of the comparative architectural schemes to obtain the architectural visual coding features corresponding to each of the comparative architectural schemes. Based on the first scale compression module, feature compression is performed on the architectural visual coding features corresponding to each of the comparative architectural schemes to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes. The first-scale compression module includes a first visual dimensionality reduction mapping component and a graph convolution processing component, wherein the graph convolution processing component includes a first spatial aggregation component and a second spatial aggregation component; based on the first-scale compression module, feature compression is performed on the architectural visual encoding features corresponding to each of the comparative architectural schemes to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes, including: Based on the first visual dimensionality reduction mapping component, the architectural visual coding features corresponding to each of the comparative architectural schemes are converted into the first visual dimensionality reduction features corresponding to each of the comparative architectural schemes. Based on the first spatial aggregation component, graph convolutional network encoding is performed on the first visual dimensionality reduction features corresponding to each of the comparative architectural schemes to obtain the first transition features corresponding to each of the comparative architectural schemes. Based on the second spatial aggregation component, graph convolutional network encoding is performed on the first transition features corresponding to each of the comparative architectural schemes to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes.

2. The method according to claim 1, characterized in that, The first-scale compression module further includes multiple sequentially cascaded first-feature compression components; the step of compressing the architectural visual coding features corresponding to each of the comparative architectural schemes based on the first-scale compression module to obtain the comparative design paradigm features corresponding to each of the comparative architectural schemes further includes: Based on the first first feature compression component among the plurality of first feature compression components, feature compression is performed on the architectural visual coding features corresponding to each of the comparative architectural schemes to obtain the first comparative design paradigm feature corresponding to each of the comparative architectural schemes. Based on the target first feature compression component among the plurality of first feature compression components, feature compression is performed on the preceding comparative design paradigm features corresponding to each of the comparative architectural schemes to obtain the target comparative design paradigm features corresponding to each of the comparative architectural schemes, wherein the preceding comparative design paradigm feature is the one preceding the target first feature compression component.

3. The method according to claim 2, characterized in that, The first architectural design generation model includes a feature integration module, which includes multiple feature integration components; The integration of the comparative design paradigm features corresponding to the multiple comparative architectural schemes yields a first integrated design paradigm feature, including: Based on the first feature integration component among the multiple feature integration components, the first comparative design paradigm features corresponding to each of the multiple comparative architectural schemes are integrated to obtain the first integrated design paradigm feature; Based on the target feature integration component among the multiple feature integration components, the target comparative design paradigm features corresponding to each of the multiple comparative architectural schemes are integrated to obtain the target first integrated design paradigm features.

4. The method according to claim 3, characterized in that, The feature integration module further includes multiple third-space aggregation components; the step of performing graph convolutional network encoding on the first integrated design paradigm features to obtain the second integrated design paradigm features includes: Based on the first third space aggregation component among the plurality of third space aggregation components, the first first integrated design paradigm feature is processed by graph convolutional network encoding to obtain the first second integrated design paradigm feature; Based on the target third space aggregation component among the plurality of third space aggregation components, the target first integrated design paradigm feature is processed by graph convolutional network encoding to obtain the target second integrated design paradigm feature.

5. The method according to claim 1, characterized in that, The method is executed based on a first architectural design generation model, which includes a spatial configuration generation module and a first feature enhancement module. The process of generating candidate architectural schemes based on the design paradigm with the second integrated design paradigm features as generation constraints and the semantic features of the requirements includes: Based on the first feature enhancement module, feature enhancement processing is performed on the demand semantic features and the second integrated design paradigm features to obtain the features of the candidate building scheme. Based on the spatial configuration generation module, the spatial configuration decoding process is performed on the features of the candidate building scheme to obtain the candidate building scheme.

6. The method according to claim 5, characterized in that, The first feature enhancement module includes a second visual dimensionality reduction mapping component and a graph convolution component, wherein the graph convolution component includes a fourth spatial aggregation component, a topological co-coding component, and a hierarchical coupling coding component; the feature enhancement processing based on the first feature enhancement module on the demand semantic features and the second integrated design paradigm features to obtain the features of the candidate architectural scheme includes: Based on the second visual dimensionality reduction mapping component, the first design parameter perturbation feature is converted into the second visual dimensionality reduction feature, and the first design parameter perturbation feature is used to describe the morphological exploration factor; Based on the fourth spatial aggregation component, the second visual dimensionality reduction feature is processed by graph convolutional network encoding to obtain the third visual dimensionality reduction feature. Based on the topological co-coding component, structural co-aggregation modeling is performed on the third visual dimensionality reduction feature and the second integrated design paradigm feature to obtain the second transition feature; Based on the hierarchical coupled coding component, structural collaborative aggregation modeling is performed on the demand semantic features and the second transition features to obtain the features of the candidate building scheme. The first architectural design generation model further includes a first scale compression module, which includes multiple sequentially cascaded first feature compression components; the first feature enhancement module includes multiple sequentially cascaded first feature expansion components; and the second integrated design paradigm feature has multiple dimensions. The step of performing feature enhancement processing on the demand semantic features and the second integrated design paradigm features based on the first feature enhancement module to obtain the features of the candidate architectural scheme further includes: Based on the first first feature extension component among the plurality of first feature extension components, feature enhancement processing is performed on the first design parameter perturbation feature, the demand semantic feature, and the first second integrated design paradigm feature to obtain the first candidate building scheme feature. The first second integrated design paradigm feature is obtained based on the feature generated by the first first feature compression component. The first design parameter perturbation feature is used to describe the morphological exploration factor. Based on the target first feature extension component among the plurality of first feature extension components, feature enhancement processing is performed on the features of the preceding candidate building schemes, the semantic features of the requirements, and the features of the target second integrated design paradigm to obtain the features of the target candidate building schemes. The features of the target second integrated design paradigm are obtained based on the features generated by the target first feature compression component.

7. The method according to claim 1, characterized in that, The method is executed based on a first architectural design generation model, the training process of which includes: Obtain architectural design requirement instances and benchmark architectural scheme instances, wherein the architectural design of the benchmark architectural scheme instances satisfies the design requirements defined by the architectural design requirement instances; Based on the second architectural design generation model, the sample requirement semantic features corresponding to the architectural design requirement instance are obtained, the first sample comparison features corresponding to the benchmark architectural scheme instance are obtained, and the second design parameter perturbation features are embedded in the first sample comparison features to obtain the second sample comparison features. The second design parameter perturbation features are used to describe the morphological exploration factor. Based on the second architectural design generation model, using the second sample comparison features as generation constraints, and based on the sample demand semantic features, generate sample candidate architectural schemes. Based on the sample candidate building schemes and the benchmark building scheme instance, the second building design generation model is trained to obtain the trained second building design generation model; Based on the trained second architectural design generation model, the first architectural design generation model is constructed. The first architectural design generation model that has completed training includes a sequentially cascaded feature extraction module, a second-scale compression module, a second-scale enhancement module, and a spatial configuration generation module. The second-scale compression module includes a sequentially cascaded first visual dimensionality reduction mapping component, a first spatial aggregation component, and a multi-scale fusion coding component. The second-scale enhancement module includes a sequentially cascaded second visual dimensionality reduction mapping component, a fourth spatial aggregation component, and a topological co-coding component. The construction of the first architectural design generation model based on the trained second architectural design generation model includes: By inheriting the first spatial aggregation component, a second spatial aggregation component is obtained; by inheriting the topology co-coding component, a hierarchical coupling coding component is obtained. By inheriting the fourth spatial aggregation component, a third spatial aggregation component is obtained; based on the feature integration component and the third spatial aggregation component, a feature integration module is constructed; In the second architectural design generation model that has completed training, the multi-scale fusion coding component of the second scale compression module is replaced with the second spatial aggregation component to obtain the first scale compression module. The hierarchical coupling coding component is embedded after the topological co-coding component of the second feature enhancement module to obtain the first feature enhancement module. The feature integration module is embedded between the first scale compression module and the first feature enhancement module to obtain the first architectural design generation model.

8. A server system, characterized in that, Includes a server, the server being used to perform the method according to any one of claims 1-7.

Images