Building scheme intelligent optimization method and system based on parametric design

Abstract

This invention relates to an intelligent optimization method for architectural schemes based on parametric design, belonging to the field of architectural design technology. Addressing the problems of existing parametric design tools, such as discrete steps, reliance on manual labor, lack of physical constraints, and insufficient consideration of constructability, this method automatically collects site, functional, regulatory, and cost parameters through multi-source interfaces to generate a set of consistent variables; it uses graph neural networks to learn topological preferences and automatically generates a compliant 3D topological skeleton; it embeds constraints from differential equations of building physical performance and outputs a Pareto optimal form solution set through gradient optimization using a differentiable surrogate model; it automatically identifies construction conflicts and generates adjustment instructions based on a constructability knowledge graph and a causal inference engine; and it introduces aesthetic constraints into the framework combined with human creative feedback to achieve human-machine collaborative evolution. This invention achieves intelligent design throughout the entire process, significantly improving efficiency, physical rationality, constructability, and creative value, and is applicable to the optimization of various types of architectural designs.

Description

Technical Field

[0001] This invention relates to the technical field of engineering design, and in particular to an intelligent optimization method and system for building schemes based on parametric design. Background Technology

[0002] With the popularization of Building Information Modeling (BIM) technology and the in-depth application of artificial intelligence in the engineering field, parametric design has become an important technical means to improve the efficiency and quality of architectural design. However, existing parametric design tools are still mainly at the level of geometric modeling assistance. The various stages of the design process (site analysis, functional layout, form optimization, construction verification) are highly discrete, relying on designers to manually integrate multi-source parameters, which is not only time-consuming and labor-intensive, but also prone to rework due to parameter conflicts and omissions.

[0003] Traditional methods rely on pre-set template libraries or random searches using genetic algorithms, resulting in fixed generation patterns that are difficult to adapt to complex and ever-changing functional requirements and site constraints, and lack innovation. The form optimization stage generally uses black-box optimization by calling external performance simulation software, which is time-consuming per iteration and the optimization results often violate the laws of building physics (such as energy conservation and lighting geometry), requiring repeated manual corrections, which is inefficient and reduces the efficiency of building scheme generation. Therefore, improvements are needed. Summary of the Invention

[0004] To improve the efficiency of architectural scheme generation, this application provides an intelligent optimization method and system for architectural schemes based on parametric design.

[0005] Firstly, the above-mentioned inventive objective of this application is achieved through the following technical solution: A method for intelligent optimization of building schemes based on parametric design, the method comprising the following steps: By acquiring site environmental parameters, functional requirement indicators, planning and regulatory constraints and cost benchmarks of building projects through multi-source design data acquisition interfaces, and generating a multi-dimensional set of parametric design variables after parameter semantic parsing and conflict detection; The set of parametric design variables is input into the intelligent topology generation engine, which learns the topology preference pattern of building space organization based on graph neural network architecture and automatically generates an initial three-dimensional topology skeleton that satisfies functional connectivity. The topological skeleton is fed into a performance-driven morphological optimization layer, which embeds the constraints of the differential equations for building physical performance simulation. The morphological parameters of the building are optimized by gradient through a differentiable performance surrogate model, and the multi-objective Pareto optimal morphological solution set that meets the performance requirements is output. The morphological solution set is structurally aligned with the constructability knowledge graph, and construction feasibility conflicts are identified through an interpretable causal reasoning engine, outputting component standardization adjustment amounts and adaptive optimization instructions for process procedures. The adjustment instructions and real-time user creative feedback are input into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints into the generative adversarial network framework to generate a family of parameterized optimization schemes that conform to human creative intentions, and outputs a comprehensive evaluation matrix of scheme performance, cost, and aesthetics.

[0006] By adopting the above technical solutions and constructing an end-to-end intelligent closed-loop architecture that extends from parameterized input, intelligent topology generation, performance-driven optimization, constructability reasoning to human-machine collaborative evolution, this solution systematically addresses the core challenges of fragmented processes, reliance on human experience, low optimization efficiency, and lack of physical rationality in traditional architectural design processes. It achieves a paradigmatic innovation in intelligent optimization of architectural solutions. Compared to existing technologies, this solution first breaks through the fragmented model of traditional manual data collection and experience-based parameter integration at the design input stage. Through automatic multi-source parameter parsing and conflict detection, it achieves global consistency verification of multi-dimensional variables such as site, function, regulations, and cost, fundamentally reducing the risk of rework due to parameter inconsistencies.

[0007] In a preferred example, this application can be further configured as follows: The step of generating a multi-dimensional parametric design variable set after obtaining site environmental parameters, functional requirement indicators, planning and regulatory constraints, and cost benchmarks for a building project through a multi-source design data acquisition interface, followed by parameter semantic parsing and conflict detection, includes the following steps: A site environment parameter parser is constructed to extract three-dimensional field data of elevation gradient, solar shading, and wind direction frequency from the digital terrain model and discretize them into learnable implicit field parameters. A functional requirement vectorization module is constructed, which transforms the building function description text into a spatial area configuration matrix, a streamline density index, and an adjacency constraint topology diagram through semantic encoding. A regulatory constraint compiler is constructed to parse the provisions on planning setbacks, height limits, and floor area ratios into a set of differentiable geometric inequalities, and embed parameter validity verification logic. A cost-sensitive parameter generator is constructed, and a correlation function of material usage, unit price, and construction difficulty is established based on the cost database to output the weight distribution of cost-driven parameters.

[0008] 3. In a preferred example, this application can be further configured such that, in the step of inputting the set of parameterized design variables into an intelligent topology generation engine, which learns topological preference patterns of architectural space organization based on a graph neural network architecture and automatically generates an initial three-dimensional topological skeleton that satisfies functional connectivity, the following steps are included: A neural network for building space organization diagrams is constructed, where nodes represent functional rooms and edges represent circulation connections. The network embeds accessibility constraints and physical rules for area allocation into the loss function. Learn typical architectural topological patterns from the built case library, capture common spatial organization features through graph isomorphic attention mechanism, and generate implicit space codes for topological patterns; The functional requirement indicators in the parameterized design variables are mapped to the latent space, and multiple candidate topology skeletons are generated by decoding. A topology diversity reward function is used to avoid pattern collapse. Perform geometric validity checks on the candidate skeletons, eliminate invalid topologies that generate self-intersections or backspace violations, and output the initial 3D topology skeleton that satisfies the basic constraints.

[0009] In a preferred example, this application can be further configured as follows: The step of feeding the topological skeleton into a performance-driven morphological optimization layer, which embeds constraints from differential equations for building physics performance simulation, performing gradient optimization of building morphological parameters through a differentiable performance surrogate model, and outputting a multi-objective Pareto optimal morphological solution set that meets performance standards includes the following steps: By embedding energy consumption simulation differential equation constraints in the building form parameter space, the calculation formulas for heat transfer of the building envelope, natural lighting coefficient, and wind pressure distribution are transformed into differentiable residual penalty terms. A multi-performance proxy model for energy consumption, lighting, and ventilation is trained using a physical information neural network architecture to ensure that its prediction results strictly satisfy the laws of energy conservation and fluid dynamics. A multi-objective gradient descent optimizer is constructed, and the weights of each performance objective are dynamically adjusted in the Pareto front search. The weight vector changes adaptively according to the cost benchmark. A form preservation regularization term is introduced to apply a Laplacian smoothing constraint to the geometric deformation during the optimization process, preventing unstructured fragmented geometry from being generated by form optimization. Specifically, the training method for the performance proxy model includes: An conformal mapping algorithm is used to transform the parameter domain of irregular building shapes into a regular computation domain, and an orthogonal sampling grid is constructed on the regular domain to improve the uniformity of training sample distribution. A physical residual term is introduced into the loss function, and a partial differential equation hard constraint is applied to the temperature field and wind speed field output by the surrogate model. After multiple iterations, the predicted field satisfies physical conservation. A transfer learning strategy was implemented, using standard building form data in the pre-training stage and injecting site-specific data of the current project in the fine-tuning stage to accelerate model convergence and improve local accuracy. An uncertainty estimation branch is integrated into the model output layer, and a performance prediction confidence interval is generated through variational inference, providing a risk quantification basis for subsequent robust optimization.

[0010] In a preferred example, this application can be further configured such that the step of structurally aligning the morphological solution set with the constructability knowledge graph, identifying construction feasibility conflicts through an interpretable causal reasoning engine, and outputting component standardization adjustment amounts and adaptive optimization instructions for process procedures includes the following steps: The component dimensions, connection methods, and hoisting space morphology are matched with the constructability knowledge graph. The knowledge graph nodes cover the standardized component library, process logic dependencies, and mechanical operation space requirements. Run the causal reasoning engine on the matching graph to identify key design parameters that cause construction conflicts, such as the size of non-standard components, process space interference, and hoisting blind spots; Based on the causal strength, the standardized adjustment amount of the component is generated, the non-standard size is mapped to the standard module as close as possible, and the alternative connection node optimization scheme is output. When a conflict involves multiple coupled processes, a process reordering instruction is automatically triggered. The system searches the knowledge graph for non-interference parallel construction sequences and outputs an adaptive optimization instruction for the process.

[0011] In a preferred example, this application can be further configured as follows: The step of inputting the adjustment instructions and real-time user creative feedback into a human-machine collaborative evolution module, which introduces pre-reviewed aesthetic preference constraints within a generative adversarial network framework to generate a family of parameterized optimized solutions that conform to the intentions of human creative thought, and outputting a comprehensive performance-cost-aesthetic evaluation matrix for the solutions, includes the following steps: A generative adversarial network for architects’ creative preferences is constructed. The discriminator learns the distribution of aesthetic features in historical schemes, while the generator samples the latent space and injects human feedback correction signals. By using the morphological solution set as the generator input, an aesthetic consistency loss is introduced into the generative adversarial training framework to ensure that the generated scheme conforms to the laws of formal beauty and style continuity. A creative feedback interaction interface is established, allowing architects to input their preferences in real time by dragging sketches and adjusting parameter sliders. The feedback signals are encoded and embedded into the generator condition vector. When outputting a family of solutions, a three-dimensional evaluation matrix of performance, cost, and aesthetics is calculated simultaneously. Each indicator adopts an interpretable scoring rule, enabling architects to understand the logic of optimization trade-offs and achieve the emergence of human-machine creative collaboration.

[0012] In a preferred example, this application can be further configured to include a parameter conflict self-detection and resolution mechanism, the specific steps of which include: Construct a hidden constraint mining tool between parameters and use association rule learning to discover non-explicit conflict patterns from the case library; Automatically scan for conflict patterns in the current parameter set. When a parameter combination is detected to trigger a conflict rule, start constraint relaxation re-optimization and adjust the parameter boundaries through a preset algorithm model. If the conflict is irreconcilable, a parameter adjustment suggestion report will be generated, pointing out the root cause of the conflict and recommending feasible parameter ranges to assist designers in conducting manual arbitration. The resolved parameter set is recorded in the conflict case library, and the rule set of the implicit constraint miner is incrementally updated to achieve continuous evolution of conflict detection capabilities.

[0013] Secondly, the above-mentioned inventive objective of this application is achieved through the following technical solutions: A smart optimization device for building schemes based on parametric design, the device comprising: a design variable set construction unit, used to acquire site environmental parameters, functional requirement indicators, planning and regulatory constraints and cost benchmarks of building projects through a multi-source design data acquisition interface, and generate a multi-dimensional parametric design variable set after parameter semantic parsing and conflict detection; A three-dimensional topology skeleton construction unit is used to input the set of parametric design variables into an intelligent topology generation engine. This engine learns the topology preference pattern of building space organization based on a graph neural network architecture and automatically generates an initial three-dimensional topology skeleton that satisfies functional connectivity. The optimal form solution set construction unit is used to feed the topological skeleton into the performance-driven form optimization layer. This layer embeds the building physics performance simulation differential equation constraints, performs gradient optimization of building form parameters through a differentiable performance proxy model, and outputs a multi-objective Pareto optimal form solution set that meets the performance standards. An adaptive optimization instruction generation unit is used to structurally align the morphological solution set with the constructability knowledge graph, identify construction feasibility conflicts through an interpretable causal reasoning engine, and output component standardization adjustment amounts and adaptive optimization instructions for process steps. The comprehensive evaluation matrix generation unit is used to input the adjustment instructions and real-time user creative feedback into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints into the generative adversarial network framework to generate a family of parameterized optimization schemes that conform to human creative intentions, and outputs a comprehensive evaluation matrix of scheme performance, cost, and aesthetics.

[0014] Thirdly, the above-mentioned objectives of this application are achieved through the following technical solutions: An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the intelligent optimization method for building schemes based on parametric design.

[0015] Fourthly, the above-mentioned objectives of this application are achieved through the following technical solutions: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent optimization method for building schemes based on parametric design. Attached Figure Description

[0016] Figure 1 This is a flowchart of an intelligent optimization method for building schemes based on parametric design, according to one embodiment of this application. Figure 2 This is a schematic diagram of a building scheme intelligent optimization device based on parametric design in one embodiment of this application; Figure 3 This is a schematic diagram of an electronic device according to an embodiment of this application.

[0017] Icon labels: 1. Design variable set construction unit; 2. Three-dimensional topology skeleton construction unit; 3. Optimal morphological solution set construction unit; 4. Adaptive optimization instruction generation unit; 5. Comprehensive evaluation matrix generation unit. Detailed Implementation

[0018] The present application will be further described in detail below with reference to the accompanying drawings.

[0019] In one embodiment, such as Figure 1 As shown, this application discloses an intelligent optimization method for building schemes based on parametric design, which specifically includes the following steps: S10: Obtain site environmental parameters, functional requirement indicators, planning and regulatory constraints and cost benchmarks for building projects through multi-source design data acquisition interfaces, and generate a multi-dimensional parametric design variable set after parameter semantic parsing and conflict detection. Specifically, taking a downtown commercial complex project as an example, the multi-source design data acquisition interface obtains raw data from four dimensions: Regarding site environmental parameters, by reading the digital terrain model to extract site elevation data, it identifies a 12-meter elevation difference between the west and east sides. Simultaneously, it obtains the surrounding building height matrix and calculates that the solar shading range of the residential building on the south side on the winter solstice is a projected length of 45 meters. Regarding functional requirements, in this embodiment, semantic parsing from the owner's submitted "Design Brief" reveals the requirement to configure 30,000 square meters of Grade A office space and 20,000 square meters of… The initial plan, consisting of "35,000 square meters of commercial retail space, a 10,000-square-meter hotel, and 600 parking spaces," was transformed into an area configuration matrix and a functional adjacency topology (office space must be adjacent to the hotel's shared meeting rooms, and commercial space must be adjacent to city roads). Regarding planning and regulatory constraints, the local planning conditions were input into a regulatory compiler, which automatically resolved them into a set of differentiable inequalities: building height limit 100 meters, floor area ratio ≤ 5.0, north setback ≥ 15 meters, and green space ratio ≥ 25%. For cost benchmarks, the system connected to the enterprise cost database and extracted a benchmark cost of 8,000 yuan / square meter for similar projects in the area. Through parameter semantic analysis and conflict detection, it was found that the initially configured 35,000 square meters of office space exceeded the floor area ratio conversion limit, triggering a conflict flag. The system automatically generated a corrected set of multi-dimensional parametric design variables, including 87 adjustable parameters (such as the area of ​​each functional block, building density, and entrance / exit orientation), providing standardized input for subsequent optimization.

[0020] S20: Input the set of parametric design variables into the intelligent topology generation engine. The engine learns the topology preference pattern of building space organization based on graph neural network architecture and automatically generates an initial three-dimensional topology skeleton that satisfies functional connectivity. The 87 design variables generated by S10 are input into the intelligent topology generation engine. Based on a graph neural network architecture, the engine learns from the spatial organization patterns of 200 excellent commercial complexes in a database of completed projects, discovering common patterns: office towers are typically located on the north side of the site to reduce western exposure and shading of the southern residential buildings; hotel lobbies need independent entrances and are close to main urban roads; and commercial podiums should be adjacent to main roads and form a "U"-shaped enclosed plaza. When the current project variables are input, the network maps office, commercial, and hotel functions as graph nodes, transforming requirements such as sunlight requirements, circulation efficiency, and noise isolation into edge weight constraints. Through a graph isomorphic attention mechanism, it decodes and generates three candidate topology skeletons: Option 1 is the classic "north tower, south podium" model; Option 2 is the innovative "twin tower sandwich" model (two office towers sandwiching a commercial atrium); and Option 3 is a "terraced" form to address sunlight constraints. A topological diversity reward function is used to avoid pattern collapse, and the final output is an initial three-dimensional topological skeleton that meets the requirements of functional connectivity and regulatory setback. The hotel entrance and exit and the commercial freight entrance are spatially separated to eliminate circulation interference.

[0021] S30: Feed the topological skeleton into the performance-driven morphological optimization layer, which embeds the differential equation constraints of building physical performance simulation, performs gradient optimization of building morphological parameters through a differentiable performance surrogate model, and outputs a multi-objective Pareto optimal morphological solution set that meets the performance standards. The topological skeleton scheme is fed into a performance-driven morphology optimization layer. This layer embeds constraints from the differential equations of building energy consumption simulation, transforming the heat transfer equations of the building envelope, the calculation formula for the natural daylighting coefficient, and the Navier-Stokes equations for wind pressure distribution into differentiable residual penalty terms. For example, when optimizing the window-to-wall ratio parameter, the residual of the heat transfer equation is automatically differentiated in each iteration. If the residual is greater than zero (violating heat conservation), the penalty term drives the parameter to adjust towards energy saving. When training the multi-performance surrogate model for energy consumption, daylighting, and ventilation, a physical information neural network architecture is used to ensure that its prediction results strictly satisfy the laws of energy conservation and fluid dynamics. After constructing a multi-objective gradient descent optimizer, the weights are dynamically adjusted in the Pareto front search: when the cost baseline is low, the weight of the energy-saving target is automatically increased; when the owner emphasizes comfort, the weight of daylighting and ventilation is increased. A morphology-preserving regularization term is introduced to prevent fragmented geometry from occurring during the optimization process. After 200 iterations of optimization, a Pareto optimal solution set containing 12 morphological solutions is output. One of the solutions achieves a balance between energy consumption (18% reduction compared to the baseline), daylighting (effective daylighting coefficient meets the standard rate of 95%), and construction cost (unit cost of 8100 yuan / square meter).

[0022] S40: Structurally align the morphological solution set with the constructability knowledge graph, identify construction feasibility conflicts through an interpretable causal reasoning engine, and output standardized component adjustment amounts and adaptive optimization instructions for process procedures. The morphological solution set was structurally aligned with the constructability knowledge graph. The knowledge graph includes nodes such as a standardized component library (beam section modulus 450×600 / 750 / 900 / 1050mm), process logic dependencies (main beam hoisting must be completed before floor slab pouring), and mechanical operation space requirements (tower crane slewing radius must cover component installation points). The causal reasoning engine identified a beam in a solution with dimensions of 450×830mm, which is a non-standard size. Causal strength analysis showed that this size required customized rebar processing, contributing a construction conflict weight of 0.65. The system automatically generated a component standardization adjustment: suggesting adjustment to the standard modulus of 450×900mm, and outputting an optimized reinforcement scheme for connection nodes. At the same time, it was detected that the clear height under the beam was only 2.1 meters, lower than the 2.5-meter working space required for tower crane hoisting, triggering a process adaptive adjustment command: the beam hoisting process was rearranged to after the floor slab formwork installation and before the rebar binding, using the formwork support to temporarily increase the working clear height and eliminate the hoisting blind spot. Through four rounds of elimination and iteration, all components met the standardization requirements, the process logic was uninterrupted, and the constructability score improved from 0.72 to 0.91.

[0023] S50: Input the adjustment instructions and real-time user creative feedback into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints in the generative adversarial network framework to generate a family of parameterized optimization schemes that conform to human creative intentions, and outputs a comprehensive evaluation matrix of scheme performance-cost-aesthetics. Architects creatively fine-tune the form solutions on an interactive platform: by dragging and dropping sketches, the top outline of the office tower was changed from a rectangle to a rounded, streamlined shape to enhance the aesthetics of the urban interface. This real-time creative feedback was encoded as a conditional vector input to the human-machine collaborative evolution module. In the generative adversarial network framework, the discriminator had learned the aesthetic feature distribution of 200 excellent historical cases. Under the constraint of rounded outlines, the generator regenerated a family of schemes: Scheme A maintained the rounded corners but added vertical lines to the curtain wall to strengthen the formal logic; Scheme B combined the rounded corners with the stepped terraces to form a spiraling shape; Scheme C only fine-tuned the rounded corner curvature to adapt to the curtain wall unit modules. The generator simultaneously calculated the performance-cost-aesthetic three-dimensional evaluation matrix for each scheme: Scheme A improved the aesthetic score by 12%, but the cost increased by 3%; Scheme B improved the aesthetic score by 20%, the cost increased by 8%, and the energy consumption factor system number increased by 5%; Scheme C improved the aesthetics by 8%, the cost increased by only 1%, and the performance remained basically unchanged. Based on the visual evaluation matrix, the architect selected scheme C as the final result, achieving a precise synergy between algorithm optimization and human creativity. It should be noted that the data is only a reference value for this embodiment of the application, and may be different data values ​​in other embodiments.

[0024] For steps S10-S50, this application systematically solves the core problems of the traditional architectural design process, such as the discrete and fragmented nature of each link, reliance on human experience, low optimization efficiency, and lack of physical rationality, by constructing an end-to-end intelligent closed-loop architecture from parameterized input, intelligent topology generation, performance-driven optimization, constructability reasoning to human-machine collaborative evolution. This achieves a paradigmatic innovation in intelligent optimization of architectural solutions.

[0025] Compared with existing technologies, this solution breaks through the fragmented mode of traditional manual data collection and experience-based parameter integration in the design input stage. Through automatic analysis and conflict detection of multi-source parameters, it achieves global consistency verification of multi-dimensional variables such as site, function, regulations, and cost, fundamentally reducing the risk of rework caused by parameter contradictions.

[0026] In the topology generation stage, this application abandons the rigid approach of relying on template library reuse or manual sketching. Instead, it innovatively adopts graph neural networks to learn the topological preference latent space of spatial organization from built cases. This can automatically generate an initial three-dimensional skeleton that meets functional connectivity and regulatory compliance, enabling topology design to move from experience imitation to data-driven intelligent creation, significantly expanding the diversity of solutions and creative possibilities.

[0027] Addressing the core bottleneck of form optimization, this application breaks through the traditional dilemma of separating performance simulation and form optimization, resulting in long iteration times and results that easily violate physical laws. By embedding the constraints of building physics differential equations such as energy consumption, lighting, and ventilation into a differentiable optimization framework, the form adjustment process naturally follows the laws of energy conservation and fluid mechanics, achieving the synergistic evolution of performance and form aesthetics, and significantly improving optimization efficiency and the physical rationality of the results.

[0028] In terms of construction feasibility verification, this application overcomes the shortcomings of traditional methods that rely on expert experience, lack systematic reasoning tools, and have difficulty in locating the root cause of conflicts. By constructing a constructability knowledge graph and an interpretable causal reasoning engine, it can automatically identify conflicts between design parameters and construction feasibility, quantify the contribution of each parameter, and output standardized adjustment amounts and process rearrangement schemes. This transforms constructability from passive inspection to proactive reasoning optimization, effectively reducing design changes during the construction phase.

[0029] Finally, this application overcomes the problem of existing algorithms being disconnected from architects' creative intentions and the optimization results failing to meet human aesthetic preferences at the human-machine collaboration level. By introducing aesthetic preference constraints and real-time creative feedback mechanisms into generative adversarial networks, it achieves a deep two-way integration of algorithmic technical rationality and architectural artistic value, enabling architects to understand the optimization trade-off logic and autonomously choose creative directions, thus completing a paradigm shift from "auxiliary tool" to "autonomous design partner".

[0030] Overall, this application has constructed a new intelligent optimization system for building solutions by systematically integrating parametric design, generative algorithms, physical information constraints and interpretable reasoning. This system has the ability to design autonomously, optimize itself, and co-create with humans and machines. It has made significant progress in improving design efficiency, ensuring physical rationality, enhancing constructability and creative value, and has pioneered a new technical path for intelligent building design.

[0031] In step S10: After acquiring site environmental parameters, functional requirement indicators, planning and regulatory constraints, and cost benchmarks for the building project through a multi-source design data acquisition interface, and generating a multi-dimensional parametric design variable set after parameter semantic parsing and conflict detection, the following steps are included: S11: Construct a site environment parameter parser to extract three-dimensional field data of elevation gradient, solar shading, and wind direction frequency from the digital terrain model and discretize them into learnable implicit field parameters. S12: Construct a functional requirement vectorization module to convert the building function description text into a spatial area configuration matrix, a streamline density index, and an adjacency constraint topology diagram through semantic encoding. S13: Construct a regulatory constraint compiler to parse the planning setback, height limit, and plot ratio clauses into a set of differentiable geometric inequalities and embed parameter validity verification logic; S14: Construct a cost-sensitive parameter generator, establish a correlation function between material usage, unit price, and construction difficulty based on the cost database, and output the weight distribution of cost-driven parameters.

[0032] For steps S11-S14, continuing with the example of the downtown commercial complex project in this application, the site environment parameter resolver extracts three-dimensional field data from the project site's digital terrain model (DTM). The resolver identifies a 12-meter elevation gradient change, with the site sloping from west to east. This gradient data is discretized into learnable implicit field parameters, enabling subsequent optimization algorithms to perceive the impact of terrain on building layout. Simultaneously, based on the surrounding building height matrix, the resolver calculates the solar shading range of the existing residential building to the south on the winter solstice, forming solar shading field parameters to ensure that the subsequently generated building form does not worsen the surrounding solar environment. For wind direction frequency, the resolver combines the annual wind rose diagram provided by the meteorological department to generate wind pressure distribution field parameters on the windward and leeward sides of the site, providing basic wind environment data for optimizing natural ventilation. This implicit representation of field parameters allows the algorithm model to directly learn the continuous spatial characteristics of the site environment, rather than relying on the coarse description of traditional discrete sampling points.

[0033] Within the same commercial complex project, the functional requirements vectorization module receives the client's design brief, which states: "30,000 square meters of Grade A office space, 20,000 square meters of retail space, 10,000 square meters of hotel space, and 600 parking spaces are required." The module uses semantic encoding technology to transform the text into a standardized spatial area configuration matrix, clearly defining the area ratio of the three functional blocks—office, retail, and hotel—as 3:2:1. Simultaneously, the module extracts implicit semantics from the text, such as "Grade A office space needs to be adjacent to the hotel to share high-end conference facilities" and "retail space needs to be adjacent to the city's main roads to enhance accessibility," generating a functional flow density index matrix to quantify the adjacency priority between office / hotel and retail / road connections. The final output is an adjacency constraint topology graph, where nodes represent functional blocks and edge weights represent the intensity of adjacency requirements. This provides directly calculable functional logic constraints for intelligent topology generation, avoiding the subjectivity and omission risks of traditional manual interpretation of requirements.

[0034] For the planning conditions of the project site, the regulatory constraint compiler automatically parses textual clauses such as "building setback from the north side of the land boundary line shall not be less than 15 meters," "building height limit of 100 meters," and "plot ratio not exceeding 5.0" into a system of differentiable geometric inequalities. For example, "setback ≥ 15 meters" is transformed into a function of the distance from the building's base coordinates to the land boundary line, and its value minus 15 meters is used as a constraint residual term; "plot ratio ≤ 5.0" is transformed into a constraint on the difference between the ratio of the total building area and the land area and 5.0. These inequalities are embedded in parameter validity verification logic. When the optimization algorithm generates building block parameters, the automatic differentiation mechanism calculates the constraint residuals during backpropagation. If the residual is positive (violating the constraint), the penalty term drives the parameters to adjust towards the valid domain. This compilation process transforms the traditional work of repeated manual verification into automatically executable calculation logic, ensuring that each generated combination of scheme variables meets the mandatory regulatory requirements and avoiding the risk of design violations.

[0035] The cost-sensitive parameter generator accesses the engineering company's cost database and discovers that similar commercial complexes in the project's location typically use reinforced concrete frame-core tube structures, with a standard floor cost of approximately 8,000-8,500 yuan per square meter. The generator establishes a correlation function between material usage, unit price, and construction difficulty: when the structural span in the building form parameters exceeds 15 meters, the correlation function automatically increases the weight of steel structure usage and the unit price increase coefficient; when the floor height exceeds 4.5 meters, it automatically increases the formwork support difficulty coefficient. Based on this function, the generator outputs a cost-driven parameter weight distribution. For example, in the optimization objective, a high cost-sensitivity weight is assigned to the structural span parameter, allowing it to actively control the span during form optimization to avoid exceeding cost limits; a medium weight is assigned to the curtain wall glass area parameter to balance lighting performance and material costs. This pre-embedded cost mechanism ensures the design scheme is economically reasonable from the initial generation stage, changing the passive model of cost accounting after the traditional design is completed.

[0036] In step S20: Inputting the set of parametric design variables into the intelligent topology generation engine, which learns the topological preference patterns of architectural space organization based on a graph neural network architecture and automatically generates an initial three-dimensional topological skeleton that satisfies functional connectivity, the steps include the following: S21: Construct a neural network for building space organization diagrams, where nodes represent functional rooms and edges represent circulation connections. The network embeds accessibility constraints and physical rules for area allocation into the loss function. Specifically, in commercial complex projects, the intelligent topology generation engine first constructs a neural network for architectural space organization. The four functional blocks—office, commercial, hotel, and parking—are abstracted as four nodes, with node attributes including area requirements and spatial characteristics. The flow requirements between functions such as "office-hotel shared meeting flow," "commercial-city road connection," and "hotel-independent drop-off area" are abstracted as weighted edges, with weight values ​​determined by the flow density index generated by S12. Reachability constraints are embedded in the network loss function, such as ensuring the path length from the office node to the hotel node does not exceed three topological hops, and physical rules for area allocation, meaning the ratio of office node area to commercial node area must meet a 3:2 configuration requirement. This graph neural network architecture embeds functional logic and planning rationality into the topology generation process from the outset, avoiding the generation of invalid connections.

[0037] S22: Learn typical building topology patterns from the built case library, capture common features of spatial organization through graph isomorphic attention mechanism, and generate latent space encoding of topology patterns; Specifically, the engine learns typical commercial complex topological patterns from a database of established case studies. This database contains 200 outstanding domestic and international projects, and a graph isomorphic attention mechanism automatically captures common features: for example, in 80% of the cases, the office tower is located on the north side of the site to reduce sunlight obstruction to the residential buildings on the south side; in 70% of the cases, hotel entrances and commercial goods entrances are spatially separated to avoid circulation intersections. The resulting latent space encoding of the topological patterns forms a low-dimensional vector space, where each vector represents a spatial organization prototype (such as the "north tower, south skirt" type, or the "twin towers sandwich" type). This encoding process transforms discrete case knowledge into a computable and interpolable continuous latent space, providing a pattern foundation for subsequent personalized generation.

[0038] S23: Map the functional requirement indicators in the parameterized design variables to the latent space, decode and generate multiple candidate topology skeletons, and use the topology diversity reward function to avoid pattern collapse; Specifically, after mapping the functional requirements indicators generated by S10 (30,000 sq m of office space, 20,000 sq m of commercial space, and 10,000 sq m of hotel space) to the hidden space, the decoder generates three candidate topological skeletons: Option 1 is the traditional "north office tower + south commercial podium" model; Option 2 is the innovative "double office towers flanking a commercial atrium" model; and Option 3 is a "terraced form" to adapt to sunlight constraints. To avoid generating repetitive patterns, the topological diversity reward function calculates the graph editing distance between the three options. If the distance is too small, the generator is penalized, forcing it to explore differentiated topological structures. The final output candidate skeletons exhibit morphological diversity based on functional connectivity (all satisfying shared meeting flow) and spatial efficiency (all meeting the floor area ratio requirements), providing architects with multiple creative starting points.

[0039] S24: Perform geometric validity checks on the candidate skeleton, eliminate invalid topologies that generate self-intersections and backspace violations, and output the initial 3D topology skeleton that satisfies the basic constraints; Specifically, during the geometric validity check of the multi-tower topology skeleton generated by candidate scheme one, the verification engine detected a geometric self-intersection at the connection between the commercial podium and the office tower (the tower core tube inserted into the podium roof, causing spatial overlap). It also found that the north outline of the tower was only 12 meters from the land boundary line, violating the regulatory requirement of a setback of ≥15 meters. The system automatically removed this invalid topology and fed back the self-intersection location and insufficient setback to the generator. After three rounds of regeneration, an initial 3D topology skeleton satisfying all geometric validity requirements (no self-intersection, compliant setback, and sufficient fire access surface) was output. The layout of the office tower, hotel tower, and commercial podium was clear, with separated circulation lines and reasonable spatial allocation, laying a good foundation for subsequent morphological optimization.

[0040] The intelligent topology generation engine, through graph neural networks and latent space learning mechanisms, achieves automated intelligent generation from functional requirements to topological skeletons, solving the pain points of traditional methods that rely on template libraries leading to rigid patterns and rely on manual sketches leading to inefficiency. Case learning through graph isomorphic attention mechanisms ensures that the generated topologies possess practically validated spatial organization rationality, avoiding infeasible solutions generated purely by algorithms. The topology diversity reward function effectively overcomes the pattern collapse problem that easily occurs in generative models, ensuring the innovativeness and richness of choices. Geometric legality verification, as the final quality checkpoint, ensures that the output topological skeleton meets all hard regulations and physical space constraints, preventing invalid designs from entering subsequent processes. Overall, this engine elevates topology design from experience-driven to a combination of data-driven and rule-driven approaches, significantly shortening the solution conception cycle and providing a high-quality starting point for subsequent refined optimization.

[0041] In step S30: Feeding the topological skeleton into a performance-driven morphology optimization layer, which embeds constraints from the differential equations of building physics performance simulation, performs gradient optimization of building morphology parameters through a differentiable performance surrogate model, and outputs a multi-objective Pareto optimal morphology solution set that meets performance standards, the steps include the following: S31: Embed the energy consumption simulation differential equation constraint in the building form parameter space, and transform the calculation formulas of heat transfer of building envelope, natural lighting coefficient and wind pressure distribution into differentiable residual penalty terms; S32: Training a multi-performance proxy model for energy consumption, lighting, and ventilation, using a physical information neural network architecture to ensure that its prediction results strictly satisfy the laws of energy conservation and fluid dynamics. S33: Construct a multi-objective gradient descent optimizer to dynamically adjust the weights of each performance objective in the Pareto front search, with the weight vector adaptively changing according to the cost benchmark; S34: Introduce a shape preservation regularization term to apply a Laplace smoothing constraint to the geometric deformation during the optimization process, preventing unstructured fragmented geometry from being generated by shape optimization. For steps S31-S34, the following are examples of the steps: Example of step S31: In the process of optimizing the form of a commercial complex, when the window-to-wall ratio parameter in the building form parameter space is set to 0.4, the optimizer calculates the residual terms of the heat transfer differential equation constraint of the building envelope. The heat transfer equation calculation shows that the winter heat loss rate under this window-to-wall ratio exceeds the energy-saving standard requirements, the residual is positive, and the penalty term immediately increases the loss function value, driving the parameter to adjust towards 0.35. At the same time, the natural daylight coefficient calculation constraint evaluates that the daylight coefficient of the first-floor commercial space under a window-to-wall ratio of 0.35 is 1.8%, which meets the daylighting standard for commercial spaces, the residual is zero, and the penalty term does not take effect. The wind pressure distribution constraint calculates the wind pressure coefficient on the windward side of the building. If the form adjustment causes the local wind pressure concentration coefficient to exceed 1.5, the residual drives the curvature adjustment of the form surface to disperse the wind pressure. These differentiable residual penalty terms directly embed the physical performance requirements into the gradient propagation path, so that each step of the form optimization automatically satisfies the basic laws of building physics.

[0042] Example of step S32: When training the energy consumption performance proxy model, a physical information neural network architecture is used. The model input consists of building morphology parameters (shape factor, window-to-wall ratio, orientation angle), and the output is the predicted annual heating and air conditioning energy consumption. A hard constraint of energy conservation is embedded in the loss function: the annual energy consumption value predicted by the model must be equal to the sum of the heat transfer losses of each building envelope component, the fresh air load, and the internal heat gain, etc. The difference is added to the total loss as a physical residual term. After 300 rounds of iterative training, the residual between the model's predicted energy consumption value and the sum of the individual components approaches zero, ensuring that the prediction results strictly satisfy the first law of thermodynamics. Similarly, when the daylighting proxy model predicts the indoor illuminance distribution, the solid angle projection law constraint is embedded; when the ventilation proxy model predicts the indoor wind speed field, the mass conservation equation constraint is embedded, ensuring that all performance predictions have physical self-consistency and avoiding prediction results that violate physical common sense that might occur with purely data-driven models.

[0043] Example of step S33: When the multi-objective gradient descent optimizer searches the Pareto front, the initial cost baseline is low. The optimizer adaptively increases the weight of the energy consumption objective, prioritizing the adjustment of morphological parameters towards energy saving (such as reducing the window-to-wall ratio and increasing the thickness of the insulation layer). As optimization progresses, the energy-saving potential of the morphology approaches its limit. The optimizer detects that further reduction in energy consumption will lead to drastic changes in morphology and loss of functional area. Therefore, it automatically reduces the weight of energy consumption and increases the weight of lighting and ventilation, shifting the focus to optimizing natural lighting and indoor airflow. The adaptive change of the weight vector dynamically balances the various objectives in the search process. The final Pareto optimal morphological solution set includes morphologies that prioritize energy saving, morphologies that prioritize lighting, and balanced morphologies. Architects can choose according to the project positioning, avoiding a single optimization bias caused by fixed weights.

[0044] Example of step S34: During morphological optimization, when the window-to-wall ratio parameter is adjusted from 0.5 to 0.35, the optimizer might divide the windows into dozens of small fragments to pursue better lighting performance, resulting in a fragmented facade morphology. By introducing a morphological preservation regularization term, the Laplace smoothing constraint calculates the rate of curvature change of the geometric deformation of the wall between windows, penalizing drastic fluctuations in the spacing between adjacent windows. For example, when the window spacing abruptly changes from 3 meters to 0.5 meters, the regularization term significantly increases the loss function value, forcing the optimizer to maintain a continuous and gradual change in window spacing. The resulting facade windows exhibit a rhythmic distribution, satisfying performance requirements while maintaining the integrity and aesthetic order of the morphology, preventing fragmented results that sacrifice morphological rationality for performance.

[0045] Specifically, the training method for the performance proxy model includes: S321: The conformal mapping algorithm is used to transform the parameter domain of irregular building form to the regular computation domain, and an orthogonal sampling grid is constructed on the regular domain to improve the uniformity of training sample distribution; S322: Introduce a physical residual term into the loss function to apply hard constraints of partial differential equations to the temperature field and wind speed field output by the surrogate model, and make the predicted field satisfy physical conservation after multiple iterations; S323: Implement a transfer learning strategy. In the pre-training stage, standard building form data is used, and in the fine-tuning stage, site-specific data of the current project is injected to accelerate model convergence and improve local accuracy. S324: Integrate the uncertainty estimation branch into the model output layer, and generate performance prediction confidence intervals through variational inference, providing a risk quantification basis for subsequent robust optimization.

[0046] In this embodiment, step S321 is illustrated as follows: The architectural form parameter domain of the commercial complex is L-shaped due to the irregular site conditions. Direct sampling in this domain would cause samples to cluster in the corners. A conformal mapping algorithm is used to transform the L-shaped parameter domain into a regular rectangular computational domain. A 5×5 orthogonal sampling grid is constructed on the rectangular domain to uniformly cover all parameter combinations. After mapping back to the L-shaped domain, sample points are automatically distributed at the site boundaries and key internal areas, avoiding sampling blind spots. This strategy significantly improves the uniformity of training sample distribution, allowing the performance proxy model to accurately predict energy consumption and lighting performance even in the apex region of the L-shaped site, avoiding local prediction distortion caused by uneven sample distribution.

[0047] Example of S322 steps: When training a surrogate model for ventilation performance, the model predicts the wind speed field distribution in an indoor atrium. A physical residual term from the Navier-Stokes equations is introduced into the loss function: the velocity divergence (∂u / ∂x + ∂v / ∂y) at each prediction point is calculated. If the divergence is not zero, mass conservation is violated. Initially, the divergence residual is large, and the physical residual term drives the adjustment of model parameters. After 200 iterations, the divergence residual approaches zero, and the predicted velocity field satisfies the mass conservation of incompressible fluids. Boundary layer velocity constraints are also introduced to ensure that the velocity gradient at the wall conforms to the laws of viscous fluid dynamics. These hard constraints ensure that the wind speed field predicted by the surrogate model is not only numerically accurate but also has smooth vector streamlines and no source sinks, fully conforming to the physical laws of fluid dynamics and can be directly used for natural ventilation design.

[0048] Example of S323 steps: The energy consumption proxy model training adopts a transfer learning strategy. In the pre-training phase, data on the morphology of 300 standard rectangular office buildings nationwide are used, allowing the model to learn the general mapping relationship between window-to-wall ratio and energy consumption. In the fine-tuning phase, site-specific data from the current commercial complex project (L-shaped plan, west-high-east-low terrain) are injected. Based on the pre-training weights, only a few parameters in the output layer and those near the input layer are adjusted, enabling the model to quickly adapt to the project characteristics. Transfer learning shortens the training time, and because the general physical laws have been mastered in the pre-training phase, the model's prediction accuracy under site-specific conditions is significantly improved after fine-tuning, especially in predicting the corner thermal bridging effect caused by the L-shaped plan.

[0049] Example of step S324: The output layer of the energy consumption proxy model integrates an uncertainty estimation branch. For a given combination of morphological parameters, the main branch predicts the annual energy consumption as a baseline value, while the variational inference branch generates a 95% confidence interval by sampling latent variables. When the morphological parameters are in a region with dense training data, the confidence interval is narrow, indicating reliable prediction; when the parameters are close to innovative forms (such as a large cantilevered irregular atrium), the confidence interval widens significantly, suggesting increased prediction uncertainty. After this interval information is passed to the morphological optimizer, the optimizer actively increases exploration sampling in high-uncertainty regions, avoiding over-searching in unreliable prediction regions, improving robust optimization capabilities, and ensuring that the final morphological solution meets performance targets while having low implementation risk.

[0050] In step S40: Structurally aligning the morphological solution set with the constructability knowledge graph, identifying construction feasibility conflicts through an interpretable causal reasoning engine, and outputting standardized component adjustment amounts and adaptive optimization instructions for process procedures, the steps include the following: S41: Match the component dimensions, connection methods, hoisting space morphology with the constructability knowledge graph. The knowledge graph nodes cover the standardized component library, process logic dependencies, and mechanical operation space requirements. S42: Run the causal reasoning engine on the matching graph to identify key design parameters that cause construction conflicts, such as non-standard component dimensions, process space interference, and hoisting blind spots; S43: Based on the causal strength, generate standardized adjustment amount of components, map non-standard dimensions to the standard module as close as possible, and output alternative connection node optimization schemes; S44: When a conflict involves multiple coupled processes, the process reordering instruction is automatically triggered. The knowledge graph searches for non-interference parallel construction sequences and outputs adaptive optimization instructions for the process.

[0051] In this embodiment of the application, in a commercial complex project, a certain scheme output by the morphology optimization layer includes a non-standard concrete beam with dimensions of 450×830mm. The subgraph matching process of the constructability knowledge graph compares the beam's size node with the "standardized component library" node in the graph, finding that the standard module library only has 450×600 / 750 / 900 / 1050mm specifications, while 830mm is a non-standard size. Simultaneously, the matching process detects a geometric overlap between the beam's hoisting path and the designed curtain wall keel installation space, creating a conflict between the hoisting space requirement node and the existing space node. Furthermore, the connection method node indicates that the connection between the beam and the core tube requires specially made embedded parts, while the "process logic dependency" node indicates that the procurement cycle for these specially made embedded parts is as long as 45 days, exceeding the normal construction cycle. Through subgraph matching, the system automatically identifies constructability conflicts in the three dimensions of size standardization, hoisting space, and procurement cycle for this beam.

[0052] When the causal inference engine is run on the matching graph, the algorithm calculates the conditional mutual information of each design parameter on construction conflicts. The results show that the "beam span" parameter contributes 0.72 to the interference in the hoisting space, making it the primary cause of the conflict; the "beam cross-section height" parameter contributes 0.68 to the cost of non-standard components, making it a key factor in cost overruns; and the "connection node type" parameter contributes 0.55 to the procurement cycle, a secondary cause of project delays. The causal inference engine generates a visualized causal link diagram, clearly showing the transmission path from "15-meter beam span" to "insufficient hoisting radius" to "interference with the curtain wall space," and the quantitative relationship from "830mm beam cross-section height" to "non-standard module" to "15% cost increase." This result transforms the ambiguous phenomenon of "construction conflicts" into clear parameter-based causes, providing a basis for precise adjustments.

[0053] Based on causal strength analysis, the system generates standardized adjustment amounts for components: mapping the beam section height of 830mm to a standard module of 900mm, and automatically adjusting the beam reinforcement ratio from 1.2% to 1.35% to meet load-bearing requirements. In the connection node optimization scheme, the system changes rigid connections that originally required specially made embedded parts to conventional hinged connections, achieving equivalent connections by increasing structural reinforcement at the beam ends, thus shortening the procurement cycle from 45 days to 7 days. For hoisting space conflicts, the system outputs two alternative solutions: Solution 1 involves prefabricating beam segments and assembling them on-site, each segment being 7.5 meters long to meet existing hoisting space requirements; Solution 2 involves adjusting the installation sequence of the curtain wall keel, installing the keel in the conflicting area after the beam hoisting is completed, thus achieving spatial temporal staggering. Architects can choose to adopt Solution 1 to maintain component integrity or Solution 2 to reduce construction complexity through the interactive interface.

[0054] When causal reasoning detects spatial interference between the "beam hoisting" and "curtain wall installation" processes, involving multiple trades, the system automatically triggers a process rescheduling instruction. Searching the constructability knowledge graph for interference-free parallel construction sequences, it is found that the "beam hoisting" process can be moved forward to immediately after the "main structure topping out," while the "curtain wall installation" process can be postponed to 3 days after the completion of the "beam hoisting." Alternative tasks that can be performed concurrently are then identified in the knowledge graph to fill this 3-day window, such as arranging "interior masonry" and "mechanical and electrical pipeline support installation" to be done concurrently. The optimized construction sequence shows that beam hoisting and curtain wall installation no longer conflict spatially, and the overall construction period only increases by 0.5 days (the window period fills the original waiting time), achieving the dual goals of interference resolution and minimizing construction time loss.

[0055] The deep integration of constructability knowledge graph and causal reasoning engine enables a paradigm shift from traditional "collision detection + manual coordination" to "intelligent recognition + causal optimization." The subgraph matching mechanism systematically compares design schemes with standardized component libraries, process logic, and spatial requirements, addressing the pain points of manual verification being prone to omissions and inefficient. The causal reasoning engine accurately locates the root causes of conflicts by quantifying contribution, transforming ambiguous construction problems into interpretable parameter causal chains, avoiding blind trial-and-error adjustments. The generation of standardized component adjustment quantities and the output of adaptive process optimization instructions achieve a fully automated closed loop from problem diagnosis to solution, elevating design-construction collaboration from passive response to proactive prevention. The multi-process coupling and rearrangement mechanism searches for feasible sequences in the knowledge graph, ensuring the global optimality of process optimization. Overall, this significantly reduces the design change rate during the construction phase, shortens the construction preparation cycle, and improves the feasibility and economy of the architectural scheme.

[0056] In step S50: The adjustment instructions and real-time user creative feedback are input into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints into the generative adversarial network framework to generate a family of parameterized optimized solutions that conform to the intention of human creativity, and outputs a comprehensive performance-cost-aesthetic evaluation matrix of the solutions. This step includes the following steps: S51: Construct an adversarial network to generate architects’ creative preferences. The discriminator learns the distribution of aesthetic features in historical schemes, and the generator samples the latent space and injects artificial feedback correction signals. S52: Using the morphological solution set as the generator input, introduce aesthetic consistency loss into the generative adversarial training framework to ensure that the generated scheme conforms to the laws of formal beauty and style continuity. S53: Establish a creative feedback interaction interface, where architects can input preferences in real time by dragging sketches and adjusting parameter sliders. The feedback signals are encoded and embedded into the generator condition vector. S54: When outputting a family of schemes, a three-dimensional evaluation matrix of performance, cost, and aesthetics is calculated simultaneously. Each indicator adopts an interpretable scoring rule, enabling architects to understand the logic of optimization trade-offs and achieve the emergence of human-machine creative collaboration.

[0057] In the optimization of commercial complex projects in this application embodiment, the human-machine collaborative evolution module first constructs an adversarial network for generating architects' creative preferences. The discriminator learns the distribution of aesthetic features from the company's historical completed design library (such as 10 modern style and 8 neoclassical style projects managed by Tiandong Construction), extracting feature patterns such as large-area glass curtain walls, simple geometric blocks, and horizontal line divisions corresponding to the "modern style." The generator initially outputs a morphological solution set, in which the office tower is a square block. At this time, the architect provides feedback through the interactive interface, "I hope the top of the tower will adopt a curved tapering shape." This human feedback is encoded as a correction signal and injected into the generator's potential space. When the discriminator evaluates the new design output by the generator, it not only judges whether it is realistic, but also whether it conforms to the learned "modern style" distribution, thereby driving the generator to understand and internalize the architect's preference for curved shapes, establishing an aesthetically pleasing generative foundation for subsequent design evolution.

[0058] S52 Step Example After inputting the morphological solution set (including a square office tower and a rectangular commercial podium) output by S30 into the generator, an aesthetic consistency loss is introduced into the adversarial training framework. This loss includes the "master-slave relationship" constraint (the tower is the master, the podium is the slave, and the volume ratio must be greater than 3:1) and the "rhythmic unity" constraint (the curtain wall segmentation module must be proportional to the floor height) from the principles of formal beauty. When the generator attempts to change the top of the office tower to a curved taper, the aesthetic consistency loss calculates the number of abrupt changes in the curve curvature. If the curvature change is too drastic, the generator is penalized to ensure a smooth curve transition and coordination with the geometric logic of the tower's main body. Through adversarial training, in the generator's output scheme family, the curved taper at the top of the tower naturally blends with the overall square block, and the curtain wall segmentation of the commercial podium also adaptively adjusts to the same horizontal line motif as the tower, avoiding stylistic breaks and ensuring the overall aesthetic quality and stylistic continuity of the scheme.

[0059] S53 Step Example Architects can adjust their designs in real time through a creative feedback interface. In a visual canvas, an architect drags the top outline of the office tower, adjusting the curve curvature from 0.1 to 0.15. The generator responds immediately, rendering the updated tower shape in real time. Simultaneously, the architect uses parameter sliders to lower the "curve taper starting height" from 80% of the tower height to 70%, and increase the "tapering range" from 5% to 8%. These feedback signals are encoded and embedded into the generator's conditional vector. Based on the original shape solution set and combined with aesthetic preference constraints, the generator quickly generates five scheme variations that meet the new constraints (such as different combinations of curves and tapering). This interface allows architects to guide the algorithm in real time through intuitive operations without writing code or adjusting complex parameters, greatly lowering the barrier to entry for intelligent design tools and achieving "what you think is what you get" creative expression.

[0060] After the generator outputs a family of design schemes, the system simultaneously calculates a three-dimensional evaluation matrix of performance, cost, and aesthetics. For scheme variant A (curve reduction of 8%), performance indicators show that energy consumption increases by 2% compared to the baseline scheme due to the increased shape coefficient, but the natural lighting coefficient increases by 1.5% due to the facade change. Cost indicators show that the curved curtain wall units need to be customized, increasing the unit cost by 80 yuan / square meter and the total cost by approximately 2.4 million yuan. Aesthetic indicators, based on discriminator scoring, show that the curved shape enhances the scheme's modernity by 15% and its harmony with the city skyline by 12%. The matrix uses interpretable scoring rules; for example, the 2.4 million yuan cost increase is interpreted as "accounting for 2.9% of the total cost, which can be reduced to 1.5% by optimizing the curtain wall module." By reading the matrix, the architect clearly understands the performance costs, cost increases, and aesthetic benefits brought by the curve reduction and proactively chooses scheme B with a "5% reduction," as its cost only increases by 900,000 yuan while still achieving an aesthetic benefit of 10%.

[0061] In summary, the human-machine collaborative evolution module, by constructing a creative preference generative adversarial network, solves the key challenge of traditional algorithm optimization in learning and internalizing architects' aesthetic styles, thus enabling the generated solutions to possess professional-grade aesthetic quality. The introduction of aesthetic consistency loss ensures that the solution family maintains stylistic unity and adherence to formal beauty principles during evolution, avoiding fragmented and distorted geometry resulting from performance optimization. The creative feedback interaction interface provides architects with intuitive and real-time guidance on algorithm generation, transforming complex parameter adjustments into low-barrier interactions such as sketch dragging and slider operations, significantly improving human-machine collaboration efficiency and user experience. The synchronous output of a three-dimensional interpretable evaluation matrix allows architects to transparently understand the performance, cost, and aesthetic impact of each decision, avoiding the dilemma of traditional black-box optimization that "only provides results without explanations," and enhancing the credibility and acceptability of algorithmic decisions.

[0062] This application embodiment also includes a parameter conflict self-detection and resolution mechanism, the specific steps of which include: S601: Construct a hidden constraint mining tool between parameters and use association rule learning to discover non-explicit conflict patterns from the case library; S602: Automatically scan for conflict patterns in the current parameter set. When a parameter combination is detected to trigger a conflict rule, start constraint relaxation re-optimization and adjust the parameter boundaries through a preset algorithm model. S603: If the conflict is irreconcilable, generate a parameter adjustment suggestion report, pointing out the root cause of the conflict and recommending feasible parameter ranges to assist designers in manual arbitration; S604: Record the resolved parameter set to the conflict case library, incrementally update the rule set of the implicit constraint miner, and realize the continuous evolution of conflict detection capability.

[0063] Specifically, in the initial parametric design phase of commercial complex projects, the implicit constraint mining algorithm initiates association rule learning from the company's database of 200 completed projects over the years. The algorithm discovers a non-explicit conflict pattern: when the three parameters "building shape coefficient greater than 0.3," "glass curtain wall window-to-wall ratio greater than 0.6," and "the average annual temperature of the project location is below 10℃" are combined, 98% of the projects in the database show cost overruns due to thickened insulation layers during the construction phase. The mining algorithm abstracts this pattern into an implicit conflict rule: high-energy-consuming structures with high window-to-wall ratios in frigid regions pose a risk of cost overruns. This rule is encoded as a computable association rule, with a confidence level of 0.92 and a lift of 3.5 recorded, providing prior knowledge for subsequent automatic scanning. This mechanism overcomes the limitations of traditional conflict identification relying on designer experience, achieving intelligent mining of implicit contradictions from historical data.

[0064] The current set of parameters for the commercial complex project was input into the conflict scanner: the average annual temperature at the project site is 8℃ (triggering low-temperature conditions), the optimized morphological solution set has a shape coefficient of 0.32 (exceeding the standard), and a window-to-wall ratio of 0.65 (exceeding the standard). The scanner detected that this parameter combination perfectly matched the conflict rules mined by S601 and immediately initiated constraint relaxation re-optimization. The system added a constraint relaxation term to the original optimization objective using the Lagrange multiplier method, automatically adjusting the parameter boundaries: dynamically tightening the upper limit of the window-to-wall ratio from 0.65 to 0.45, while relaxing the optimization objective of the shape coefficient from 0.32 to 0.35, and introducing an insulation layer cost coefficient as a penalty term. The new morphological solution set generated after re-optimization had a window-to-wall ratio reduced to 0.42 and a shape coefficient optimized to 0.34. Although energy consumption increased slightly, the insulation layer cost decreased by 15%, and the overall cost remained within a controllable range, achieving intelligent conflict resolution and multi-objective rebalancing.

[0065] When the scanner detected an irreconcilable conflict between "the hotel lobby's clear height needs to be 6 meters" and "the structural beam height is limited to within 0.8 meters" (due to the 18-meter span requiring a beam height of at least 1.2 meters), conflict resolution failed. The system automatically generated a parameter adjustment suggestion report, clearly pointing out the root cause of the conflict: "The hotel lobby's 18-meter span requires a beam height ≥ 1.2 meters, which contradicts the owner's insistence on a 0.8-meter beam height limit." The report recommended two feasible parameter ranges: Option 1 reduces the lobby span to 12 meters, lowering the beam height to 0.75 meters, requiring a reduction of 300 square meters in lobby area; Option 2 uses prestressed beam technology, compressing the beam height to 0.85 meters for the 18-meter span, but requiring an additional 1.2 million yuan in prestressing-specific costs. The report uses visual charts to illustrate the impact weight of each option on function, cost, and construction period, helping designers quickly understand the advantages and disadvantages. After manual arbitration, Option 2 was selected. After the designer confirmed on the interactive interface, the parameter set was automatically updated, and the system continued the optimization process.

[0066] After successfully resolving the "high window-to-wall ratio in frigid regions" conflict, the system records the final parameter set (window-to-wall ratio 0.42, shape coefficient 0.34, and adjustment amount for insulation layer cost coefficient) in the conflict case database and marks the case as "Hidden Conflict Discovery - Automatic Resolution Successful." The association rule learning module of the hidden constraint mining tool incrementally updates the database, increasing the number of collaborative samples for this case from 200 to 201 and fine-tuning the rule confidence to 0.93. During the next scan of a new project, the trigger threshold and penalty intensity of this rule will be automatically optimized based on the updated statistics, resulting in more accurate identification and more precise resolution of similar conflicts. This mechanism evolves the conflict detection capability from a static rule database to a continuously learning dynamic system that evolves with accumulated project experience, overcoming the shortcomings of traditional methods that suffer from fixed knowledge and an inability to self-improve.

[0067] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0068] In one embodiment, a building scheme intelligent optimization device based on parametric design is provided, which corresponds one-to-one with the building scheme intelligent optimization method based on parametric design described in the above embodiments. For example... Figure 2 As shown, the intelligent optimization device for building solutions based on parametric design includes: Design variable set construction unit 1 is used to obtain site environment parameters, functional requirement indicators, planning and regulatory constraints and cost benchmarks of building projects through multi-source design data acquisition interface, and generate a multi-dimensional parameterized design variable set after parameter semantic parsing and conflict detection. The three-dimensional topology skeleton construction unit 2 is used to input the set of parametric design variables into the intelligent topology generation engine. This engine learns the topology preference pattern of building space organization based on graph neural network architecture and automatically generates an initial three-dimensional topology skeleton that satisfies functional connectivity. The optimal form solution set construction unit 3 is used to feed the topological skeleton into the performance-driven form optimization layer. This layer embeds the building physical performance simulation differential equation constraints, performs gradient optimization of building form parameters through a differentiable performance proxy model, and outputs a multi-objective Pareto optimal form solution set that meets the performance standards. Adaptive optimization instruction generation unit 4 is used to structurally align the morphological solution set with the constructability knowledge graph, identify construction feasibility conflicts through an interpretable causal reasoning engine, and output component standardization adjustment amount and adaptive optimization instructions for process steps. The comprehensive evaluation matrix generation unit 5 is used to input the adjustment instructions and real-time user creative feedback into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints in the generative adversarial network framework to generate a family of parameterized optimization schemes that conform to human creative intentions, and outputs a comprehensive evaluation matrix of scheme performance-cost-aesthetics.

[0069] Specific limitations regarding the intelligent optimization device for building schemes based on parametric design can be found in the limitations of the intelligent optimization method for building schemes based on parametric design mentioned above, and will not be repeated here. Each module in the aforementioned intelligent optimization device for building schemes based on parametric design can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the electronic device, or stored in the memory of the electronic device as software, so that the processor can call and execute the corresponding operations of each module.

[0070] In one embodiment, an electronic device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, this electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The database stores the database. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an intelligent optimization method for building schemes based on parametric design.

[0071] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: By acquiring site environmental parameters, functional requirement indicators, planning and regulatory constraints and cost benchmarks of building projects through multi-source design data acquisition interfaces, and generating a multi-dimensional set of parametric design variables after parameter semantic parsing and conflict detection; The set of parametric design variables is input into the intelligent topology generation engine, which learns the topology preference pattern of building space organization based on graph neural network architecture and automatically generates an initial three-dimensional topology skeleton that satisfies functional connectivity. The topological skeleton is fed into a performance-driven morphological optimization layer, which embeds the constraints of the differential equations for building physical performance simulation. The morphological parameters of the building are optimized by gradient through a differentiable performance surrogate model, and the multi-objective Pareto optimal morphological solution set that meets the performance requirements is output. The morphological solution set is structurally aligned with the constructability knowledge graph, and construction feasibility conflicts are identified through an interpretable causal reasoning engine, outputting component standardization adjustment amounts and adaptive optimization instructions for process procedures. The adjustment instructions and real-time user creative feedback are input into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints into the generative adversarial network framework to generate a family of parameterized optimization schemes that conform to human creative intentions, and outputs a comprehensive evaluation matrix of scheme performance, cost, and aesthetics.

[0072] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: By acquiring site environmental parameters, functional requirement indicators, planning and regulatory constraints and cost benchmarks of building projects through multi-source design data acquisition interfaces, and generating a multi-dimensional set of parametric design variables after parameter semantic parsing and conflict detection; The set of parametric design variables is input into the intelligent topology generation engine, which learns the topology preference pattern of building space organization based on graph neural network architecture and automatically generates an initial three-dimensional topology skeleton that satisfies functional connectivity. The topological skeleton is fed into a performance-driven morphological optimization layer, which embeds the constraints of the differential equations for building physical performance simulation. The morphological parameters of the building are optimized by gradient through a differentiable performance surrogate model, and the multi-objective Pareto optimal morphological solution set that meets the performance requirements is output. The morphological solution set is structurally aligned with the constructability knowledge graph, and construction feasibility conflicts are identified through an interpretable causal reasoning engine, outputting component standardization adjustment amounts and adaptive optimization instructions for process procedures. The adjustment instructions and real-time user creative feedback are input into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints into the generative adversarial network framework to generate a family of parameterized optimization schemes that conform to human creative intentions, and outputs a comprehensive evaluation matrix of scheme performance, cost, and aesthetics.

[0073] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0074] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0075] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for intelligent optimization of building schemes based on parametric design, characterized in that, The method includes the following steps: obtaining site environmental parameters, functional requirement indicators, planning and regulatory constraints and cost benchmarks of the building project through a multi-source design data acquisition interface; and generating a multi-dimensional parametric design variable set after parameter semantic parsing and conflict detection. The set of parametric design variables is input into the intelligent topology generation engine, which learns the topology preference pattern of building space organization based on graph neural network architecture and automatically generates an initial three-dimensional topology skeleton that satisfies functional connectivity. The topological skeleton is fed into a performance-driven morphological optimization layer, which embeds the constraints of the differential equations for building physical performance simulation. The morphological parameters of the building are optimized by gradient through a differentiable performance surrogate model, and the multi-objective Pareto optimal morphological solution set that meets the performance requirements is output. The morphological solution set is structurally aligned with the constructability knowledge graph, and construction feasibility conflicts are identified through an interpretable causal reasoning engine, outputting component standardization adjustment amounts and adaptive optimization instructions for process procedures. The adjustment instructions and real-time user creative feedback are input into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints into the generative adversarial network framework to generate a family of parameterized optimization schemes that conform to human creative intentions, and outputs a comprehensive evaluation matrix of scheme performance, cost, and aesthetics.

2. The intelligent optimization method for building schemes based on parametric design according to claim 1, characterized in that, The step of generating a multi-dimensional parametric design variable set after acquiring site environmental parameters, functional requirement indicators, planning and regulatory constraints, and cost benchmarks for a building project through a multi-source design data acquisition interface, and performing parameter semantic parsing and conflict detection, includes the following steps: A site environment parameter parser is constructed to extract three-dimensional field data of elevation gradient, solar shading, and wind direction frequency from the digital terrain model and discretize them into learnable implicit field parameters. A functional requirement vectorization module is constructed, which transforms the building function description text into a spatial area configuration matrix, a streamline density index, and an adjacency constraint topology diagram through semantic encoding. A regulatory constraint compiler is constructed to parse the provisions on planning setbacks, height limits, and floor area ratios into a set of differentiable geometric inequalities, and embed parameter validity verification logic. A cost-sensitive parameter generator is constructed, and a correlation function of material usage, unit price, and construction difficulty is established based on the cost database to output the weight distribution of cost-driven parameters.

3. The intelligent optimization method for building schemes based on parametric design according to claim 1, characterized in that, The step of inputting the set of parametric design variables into the intelligent topology generation engine, which learns the topological preference patterns of architectural space organization based on a graph neural network architecture and automatically generates an initial three-dimensional topological skeleton that satisfies functional connectivity, includes the following steps: A neural network for building space organization diagrams is constructed, where nodes represent functional rooms and edges represent circulation connections. The network embeds accessibility constraints and physical rules for area allocation into the loss function. Learn typical architectural topological patterns from the built case library, capture common spatial organization features through graph isomorphic attention mechanism, and generate implicit space codes for topological patterns; The functional requirement indicators in the parameterized design variables are mapped to the latent space, and multiple candidate topology skeletons are generated by decoding. A topology diversity reward function is used to avoid pattern collapse. Perform geometric validity checks on the candidate skeletons, eliminate invalid topologies that generate self-intersections or backspace violations, and output the initial 3D topology skeleton that satisfies the basic constraints.

4. The intelligent optimization method for building schemes based on parametric design according to claim 3, characterized in that, The step of feeding the topological skeleton into a performance-driven morphological optimization layer, which embeds constraints from the differential equations of building physics performance simulation, and performing gradient optimization of building morphological parameters through a differentiable performance surrogate model to output a multi-objective Pareto optimal morphological solution set that meets performance standards, includes the following steps: By embedding energy consumption simulation differential equation constraints in the building form parameter space, the calculation formulas for heat transfer of the building envelope, natural lighting coefficient, and wind pressure distribution are transformed into differentiable residual penalty terms. A multi-performance proxy model for energy consumption, lighting, and ventilation is trained using a physical information neural network architecture to ensure that its prediction results strictly satisfy the laws of energy conservation and fluid dynamics. A multi-objective gradient descent optimizer is constructed, and the weights of each performance objective are dynamically adjusted in the Pareto front search. The weight vector changes adaptively according to the cost benchmark. A form preservation regularization term is introduced to apply a Laplacian smoothing constraint to the geometric deformation during the optimization process, preventing unstructured fragmented geometry from being generated by form optimization. Specifically, the training method for the performance proxy model includes: An conformal mapping algorithm is used to transform the parameter domain of irregular building shapes into a regular computation domain, and an orthogonal sampling grid is constructed on the regular domain to improve the uniformity of training sample distribution. A physical residual term is introduced into the loss function, and a partial differential equation hard constraint is applied to the temperature field and wind speed field output by the surrogate model. After multiple iterations, the predicted field satisfies physical conservation. A transfer learning strategy was implemented, using standard building form data in the pre-training stage and injecting site-specific data of the current project in the fine-tuning stage to accelerate model convergence and improve local accuracy. An uncertainty estimation branch is integrated into the model output layer, and a performance prediction confidence interval is generated through variational inference, providing a risk quantification basis for subsequent robust optimization.

5. The intelligent optimization method for building schemes based on parametric design according to claim 4, characterized in that, The step of aligning the morphological solution set with the constructability knowledge graph, identifying construction feasibility conflicts through an interpretable causal reasoning engine, and outputting standardized component adjustment amounts and adaptive optimization instructions for process procedures includes the following steps: The component dimensions, connection methods, and hoisting space morphology are matched with the constructability knowledge graph. The knowledge graph nodes cover the standardized component library, process logic dependencies, and mechanical operation space requirements. Run the causal reasoning engine on the matching graph to identify key design parameters that cause construction conflicts, such as the size of non-standard components, process space interference, and hoisting blind spots; Based on the causal strength, the standardized adjustment amount of the component is generated, the non-standard size is mapped to the standard module as close as possible, and the alternative connection node optimization scheme is output. When a conflict involves multiple coupled processes, a process reordering instruction is automatically triggered. The system searches the knowledge graph for non-interference parallel construction sequences and outputs an adaptive optimization instruction for the process.

6. The intelligent optimization method for building schemes based on parametric design according to claim 5, characterized in that, The step of inputting the adjustment instructions and real-time user creative feedback into the human-machine collaborative evolution module, which introduces pre-reviewed aesthetic preference constraints into the generative adversarial network framework to generate a family of parameterized optimized solutions that conform to the intentions of human creativity, and outputs a comprehensive performance-cost-aesthetic evaluation matrix of the solutions, includes the following steps: A generative adversarial network for architects’ creative preferences is constructed. The discriminator learns the distribution of aesthetic features in historical schemes, while the generator samples the latent space and injects human feedback correction signals. By using the morphological solution set as the generator input, an aesthetic consistency loss is introduced into the generative adversarial training framework to ensure that the generated scheme conforms to the laws of formal beauty and style continuity. A creative feedback interaction interface is established, allowing architects to input their preferences in real time by dragging sketches and adjusting parameter sliders. The feedback signals are encoded and embedded into the generator condition vector. When outputting a family of solutions, a three-dimensional evaluation matrix of performance, cost, and aesthetics is calculated simultaneously. Each indicator adopts an interpretable scoring rule, enabling architects to understand the logic of optimization trade-offs and achieve the emergence of human-machine creative collaboration.

7. The intelligent optimization method for building schemes based on parametric design according to claim 6, characterized in that, A parameter conflict self-detection and resolution mechanism is set up, and the specific steps include: Construct a hidden constraint mining tool between parameters and use association rule learning to discover non-explicit conflict patterns from the case library; Automatically scan for conflict patterns in the current parameter set. When a parameter combination is detected to trigger a conflict rule, start constraint relaxation re-optimization and adjust the parameter boundaries through a preset algorithm model. If the conflict is irreconcilable, a parameter adjustment suggestion report will be generated, pointing out the root cause of the conflict and recommending feasible parameter ranges to assist designers in conducting manual arbitration. The resolved parameter set is recorded in the conflict case library, and the rule set of the implicit constraint miner is incrementally updated to achieve continuous evolution of conflict detection capabilities.

8. A building scheme intelligent optimization device based on parametric design, applied to any one of the building scheme intelligent optimization methods based on parametric design as described in claims 1-7, characterized in that, The device includes: The design variable set construction unit (1) is used to obtain site environment parameters, functional requirement indicators, planning and regulatory constraints and cost benchmarks of building projects through multi-source design data acquisition interface, and generate a multi-dimensional parameterized design variable set after parameter semantic parsing and conflict detection. The three-dimensional topology skeleton construction unit (2) is used to input the set of parameterized design variables into the intelligent topology generation engine. The engine learns the topology preference pattern of building space organization based on graph neural network architecture and automatically generates an initial three-dimensional topology skeleton that satisfies functional connectivity. The optimal form solution set construction unit (3) is used to feed the topological skeleton into the performance-driven form optimization layer. This layer is embedded with the building physical performance simulation differential equation constraint. The building form parameters are optimized by gradient through a differentiable performance proxy model, and the multi-objective Pareto optimal form solution set with performance meets the standard is output. The adaptive optimization instruction generation unit (4) is used to structurally align the morphological solution set with the constructability knowledge graph, identify construction feasibility conflicts through an interpretable causal reasoning engine, and output component standardization adjustment amount and process adaptive optimization instructions. The comprehensive evaluation matrix generation unit (5) is used to input the adjustment instructions and real-time user creative feedback into the human-machine collaborative evolution module. This module introduces pre-reviewed aesthetic preference constraints in the generative adversarial network framework to generate a family of parameterized optimization schemes that conform to human creative intentions and outputs a comprehensive evaluation matrix of scheme performance-cost-aesthetics.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent optimization method for building schemes based on parametric design as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent optimization method for building schemes based on parametric design as described in any one of claims 1 to 7.

Images