Building wall intelligent design method based on semantic-physical double closed loop

Abstract

This invention provides an intelligent design method for building walls based on a semantic-physical dual closed loop, comprising the following steps: acquiring project data and environmental element data of the target project, generating semantic tags and project context, retrieving standard provisions based on semantic tags, and compiling the standard provisions into constraint expressions; constructing an executable design grammar for wall design, establishing a bidirectional mapping relationship between the executable design grammar and the building information model, and generating multiple candidate wall schemes conforming to the executable design grammar under the constraints of the constraint expressions; using a surrogate model to perform multiphysics performance evaluation on the candidate wall schemes, triggering a high-precision solver to verify based on the uncertainty evaluation results, and selecting a set of compliant schemes; iteratively optimizing the set of compliant schemes under the penalty mechanisms of hard and soft constraints, outputting a Pareto optimal solution set, parsing the Pareto optimal solution set based on the executable design grammar, and generating design documents and a compliance audit report.

Description

Technical Field

[0001] This invention belongs to the field of building intelligent design technology, specifically relating to a building wall intelligent design method based on semantic-physical dual closed loop. Background Technology

[0002] Walls are the main body of a building envelope, and their thermal and wet performance directly affects building energy consumption, indoor thermal comfort, durability, and health risks, such as condensation and mold. In the design process of building walls, the existing design process relies on manual retrieval of specifications and atlases, experience-based structural combinations, and discrete simulation tools. This traditional approach has gradually revealed many problems in practical applications.

[0003] With the development of Building Information Modeling (BIM) and artificial intelligence technologies, intelligent design methods are gradually being applied in the field of architectural design. This method integrates and optimizes various constraints and objectives in the design process by introducing technologies such as optimizers, thereby improving design efficiency and quality.

[0004] However, existing intelligent design methods have the following shortcomings: the understanding and application of specification clauses rely on manual interpretation, lacking an automated conversion mechanism from natural language specifications to executable constraints, making it difficult to guarantee the compliance of design schemes; multiphysics performance evaluation usually adopts high-precision numerical simulation, which is computationally expensive and time-consuming, making it difficult to meet the real-time requirements of design iteration optimization; there is a lack of a closed-loop feedback mechanism from the semantic level to the physical level, and the generation of design schemes and performance evaluation are disconnected, making it impossible to achieve true intelligent design optimization; the design process lacks a complete traceability link, making it difficult to meet the strict requirements of compliance audits in construction engineering. Summary of the Invention

[0005] This invention proposes an intelligent design method for building walls based on a semantic-physical dual closed loop, which solves the problem that legal provisions are not machine-readable and it is difficult to form constraint functions that can be directly called by the optimizer.

[0006] To address the aforementioned technical problems, this invention provides a smart design method for building walls based on a semantic-physical dual closed loop, comprising the following steps:

[0007] Step S1: Obtain project data and environmental element data of the target project, generate semantic tags and project context, retrieve specification clauses based on semantic tags, and compile the specification clauses into constraint expressions containing hard constraints, soft constraints and violation degree functions;

[0008] Step S2: Construct an executable design syntax for wall design, establish a bidirectional mapping relationship between the executable design syntax and the building information model, and generate multiple candidate wall schemes that conform to the executable design syntax under the constraints of the constraint expressions;

[0009] Step S3: Use the surrogate model to perform multiphysics performance evaluation on the candidate wall schemes, trigger the high-precision solver to verify the results based on the uncertainty evaluation, and select the set of compliant schemes;

[0010] Step S4: Iteratively optimize the set of compliant solutions under the penalty mechanisms of hard and soft constraints, output the Pareto optimal solution set, parse the Pareto optimal solution set based on the executable design syntax, and generate design documents and compliance audit reports.

[0011] Preferably, in step S1, the specification clauses are compiled into constraint expressions by a constraint compiler, wherein the constraint compiler includes a clause vector indexing unit, a clause parser, and a constraint compilation unit;

[0012] The text vector indexing unit performs text vectorization processing on the texts in regulations, atlases and enterprise standards and establishes an index library;

[0013] The text parser extracts the target text from the index and decomposes it into a variable set, applicable scope, and performance threshold requirements.

[0014] The constraint compilation unit constructs a dictionary for the building wall domain, performs dependency parsing on natural language texts, and extracts constraint objects, constraint relationships, thresholds, and applicable scope.

[0015] Preferably, the constraint compilation unit transforms the hard constraints into a combination of first-order predicate logic and mathematical formulas and associates them with an infinite penalty function through a logical rule mapping algorithm, and transforms the soft constraints into formal expressions and associates them with a quadratic penalty function.

[0016] Preferably, the executable design syntax in step S2 includes formal definitions of wall layer sequence, material moisture resistance characteristics, air tightness level, thermal bridge parameters, node construction, and anchor density.

[0017] Preferably, the generation of the wall candidate scheme in step S2 includes the following steps: converting the constraint expression into structured instructions through prompting engineering, calling the material genes, node genes and anchor genes in the construction gene library; using a large language model to generate a wall candidate scheme that conforms to the executable design syntax based on the retrieved gene data combination, and outputting the clause-construction-performance mapping information.

[0018] Preferably, the conditions for triggering the high-precision solver verification in step S3 include:

[0019] (1) The variance of the performance index output by the proxy model is greater than or equal to the preset variance threshold;

[0020] (2) The confidence interval width output by the proxy model is greater than or equal to a preset width threshold;

[0021] (3) The performance parameters of the candidate wall scheme are within the preset range of the hard constraint boundary.

[0022] Preferably, in step S4, a hybrid optimization strategy is used to iteratively optimize the set of compliant solutions. The hybrid optimization strategy includes a hierarchical evolutionary algorithm, a Bayesian optimization algorithm, and a policy gradient algorithm. The hierarchical evolutionary algorithm is used for global topology search, the Bayesian optimization algorithm is used for local optimization of continuous parameters, and the policy gradient algorithm is used for dynamically adjusting hyperparameters during the optimization process.

[0023] Preferably, in step S4, the target weights and penalty coefficients are updated online through a meta-learning mechanism. The meta-learning mechanism extracts meta-knowledge based on historical project data and generates initial target weights and penalty coefficients. During the iterative optimization process, the target weights and the penalty coefficients of the soft constraints are updated online based on real-time performance feedback.

[0024] Preferably, the compliance audit report in step S4 is archived in both machine-readable and document formats. The compliance audit report includes a complete reasoning chain from the source of the regulatory provisions, constraint expression transformation, performance evaluation data to the final compliance conclusion, as well as version metadata.

[0025] Preferably, step S4 further includes a model update and optimization step: during the operation phase, measured data of the building are collected, an adaptive window detection algorithm is used to handle discrete data drift, a distribution difference statistical test algorithm is used to handle continuous data drift, when data drift or performance degradation is detected, incremental data after drift is selected to construct a hybrid training set to fine-tune the proxy model, conflicts between new data and existing knowledge graphs are identified, entity attributes or inference rules in the knowledge graph are updated based on data timeliness priority and domain specification constraints, and difference annotations and re-optimization suggestions are generated.

[0026] The beneficial effects of the present invention include at least the following:

[0027] 1. This invention automatically transforms natural language specification clauses into constraint expressions containing hard constraints, soft constraints, and violation degree functions through a constraint compiler, realizing the automated conversion from semantic level to formal constraints, effectively ensuring the compliance of design schemes, and avoiding omissions and errors that may occur from manual interpretation of specifications;

[0028] 2. This invention constructs an executable design syntax for wall design and establishes a bidirectional mapping relationship with the building information model. Combined with the generation capability of a large language model, it can quickly generate multiple candidate wall schemes that conform to the syntax specifications, significantly improving design efficiency.

[0029] 3. A surrogate model is used for the initial evaluation of multiphysics performance, and a high-precision solver is conditionally triggered for verification based on the uncertainty evaluation results. This significantly reduces the computational cost while ensuring the evaluation accuracy, achieving a balance between computational efficiency and accuracy.

[0030] 4. This invention employs a hybrid optimization strategy combined with a meta-learning mechanism for iterative optimization, which can effectively combine global search and local optimization, and dynamically adjust optimization parameters based on real-time feedback, thereby improving optimization efficiency and solution quality. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0032] Figure 2 This is a flowchart illustrating the hybrid optimization strategy of an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0034] like Figure 1 As shown, this invention provides an intelligent design method for building walls based on a semantic-physical dual closed loop. This method establishes a bidirectional closed-loop feedback mechanism between the semantic and physical layers, achieving intelligent design throughout the entire process, from project requirement analysis, specification constraint compilation, scheme generation and evaluation to iterative optimization output. Specifically, it includes the following steps:

[0035] Step S1: Obtain project data and environmental element data of the target project, generate semantic tags and project context, retrieve specification clauses based on semantic tags, and compile the specification clauses into constraint expressions containing hard constraints, soft constraints and violation degree functions.

[0036] Specifically, project data includes energy consumption targets, thermal and humidity safety requirements, carbon emission limits, cost budgets, and construction method constraints. Environmental element data includes climate zones, building types, and applicable standard versions. The automatic generation of semantic tags follows a four-step logic: data input, parsing and matching, encoding generation, and contextual association, requiring no manual intervention throughout. The specific process is as follows:

[0037] Step S1-1: Structured data acquisition.

[0038] The system automatically retrieves two types of core information through pre-set standardized forms or by connecting to project management platforms and building information modeling interfaces. Project objectives include energy consumption targets, such as "≤65kWh / ( Requirements for thermal and humidity safety, such as "no condensation", and carbon emission limits, such as "≤80" kg / m ²”, cost budget, such as “≤1200 yuan / Construction methods are subject to constraints, such as "prohibition of wet work on site".

[0039] The climate zone in the environmental elements is automatically matched to the GB50176-2016 zoning standard based on the project's geographical location, such as "Severe Cold Zone A". The building type is extracted from the project's filing information, such as "Office Building". The applicable standard version is preferentially read from the version specified by the project. If no version is specified, the latest current version is used by default, such as "GB50189-2015 (2022 Edition)".

[0040] Step S1-2: Feature classification and rule matching.

[0041] The system has a built-in dual-dimensional tag rule library for objectives and environment. The objective dimension is pre-defined with five categories of tag prefixes: "Energy Consumption (TE), Thermal and Humidity Safety (TH), Carbon Emissions (TC), Cost (T-Cost), and Construction Method (TM)". Each prefix corresponds to a fixed coding rule, such as the energy consumption objective "≤65kWh / ( The system matches “TE-01” with the environmental dimension and automatically associates it with the corresponding standard number. The environmental dimension is pre-defined with three types of label prefixes: “Climate Zone (EC), Building Type (EB), and Standard Version (ES)”. For example, “Climate Zone A” matches “EC-01”, and “Office Building” matches “EB-03”. The system automatically identifies the dimension to which the collected information belongs and accurately matches it with the label prefixes and coding logic in the rule base.

[0042] Step S1-3: Automatic label encoding generation.

[0043] The final semantic tags are generated according to the format of prefix + attribute value + standard association. For example, the energy consumption target "≤65kWh / ( (Conforming to GB50189-2015) generates “TE-01_≤65kWh / ( The system generates "EC-01_Severe Cold Zone A_GB50176-2016" for the climate zone "Severe Cold Zone A (conforms to GB50176-2016)". The system automatically verifies the uniqueness of the tags. If there are the same elements, such as different items all being "Severe Cold Zone A", the tag codes are kept consistent to ensure the uniformity of subsequent searches and constraint matching.

[0044] Step S1-4: Context association binding.

[0045] All generated semantic tags are bound to basic project information to form a project-specific context. Basic project information includes project name, construction location, and construction unit. Relationships are established between tags; for example, "EC-01_Severe Cold Zone A" is automatically associated with "TE-01_≤65kWh / ( Because energy consumption standards are more stringent in frigid regions; by embedding tags into the project context database, subsequent steps can directly call the tags to quickly locate the applicable standard provisions and structural genes of the project, and realize the automatic linkage of elements-tags-design.

[0046] The semantic tags are divided into two dimensions: target and environment. They adopt a hierarchical coding + attribute value structure, with a total of 12 items. The target dimension includes 5 items such as energy consumption and thermal and humidity safety, while the environment dimension includes 7 items such as climate zone and building type.

[0047] The specification clauses are compiled into constraint expressions containing hard constraints, soft constraints, and violation degree functions by the constraint compiler. The constraint compiler consists of three core components: a clause vector indexing unit, a clause parser, and a constraint compilation unit.

[0048] The clause vector indexing unit performs text vectorization processing on clauses in regulations, atlases, and enterprise standards and establishes an index library, converting the clauses into a computer-recognizable vector form to facilitate subsequent rapid retrieval and matching.

[0049] The clause parser extracts the target clause from the index and breaks it down into the set of variables in the clause, such as the wall heat transfer coefficient U-value and the linear heat loss coefficient. Values, applicable scope, such as specific climate zones, building types, and performance threshold requirements, such as U-value ≤ 0.35W / ( At the same time, the logical relationships between the various elements within the clauses are sorted out.

[0050] The constraint compilation unit transforms the parsed results into constraint functions that can be directly called by the optimizer. This unit first constructs a dictionary for the building wall domain, including standard terminology such as "heat transfer coefficient U-value" and "combustion performance rating"; parameter units such as W / ( The algorithm uses MPa as the basis for parsing constraints, including predicates such as "≤", "should not be lower than", "must meet", and standard numbers such as GB50189-2015. It performs dependency parsing on the natural language text to extract constraint objects, constraint relationships, thresholds, and applicable scope. Then, a logical rule mapping algorithm transforms hard constraints into a combination of first-order predicate logic and mathematical formulas, associated with an infinite penalty function, while soft constraints are transformed into formal expressions and associated with a quadratic penalty function.

[0051] For hard constraints, such as "the heat transfer coefficient U-value of the exterior wall in the extremely cold zone A is ≤0.30W / ( The text describes a process involving dependency parsing to identify subject-verb-object structures and extract constraints (heat transfer coefficient U-value of the exterior wall in frigid zone A), constraint relationships (≤), and threshold values ​​(0.30 W / ().). The core elements of "applicable scope (severe cold zone A)" are transformed into a first-order predicate logic expression. Exterior wall (exterior wall, climate zone = severe cold zone A → exterior wall. U value ≤ 0.30W / ( ")" triggers an infinite penalty function when violated. .

[0052] For soft constraints, such as "the air tightness level of the external wall should not be lower than level 6", the parsing is "the air tightness level of the external wall is ≥ level 6", and a quadratic penalty function is associated:

[0053] ;

[0054] In the formula, This is the penalty coefficient.

[0055] In building wall design, clauses compiled as hard constraints are mostly mandatory code requirements; violations directly exclude the design. For example, GB50189-2015 stipulates that the heat transfer coefficient U-value of exterior walls in severely cold zone A should be ≤0.30 W / ( The mandatory safety and energy-saving clauses include GB50016-2014 (2018 edition) requiring that the combustion performance rating of external wall insulation materials be no lower than B1 when the building height is >24m, and JGJ235-2011 requiring that the tensile bearing capacity of anchors be ≥0.6kN. Clauses compiled as soft constraints are mostly non-mandatory optimization requirements, with secondary penalties for violations based on the degree of deviation. Examples include GB / T50378-2019 requiring that the airtightness level of external walls be no lower than level 6 (bonus point), JGJ144-2019 requiring that the ambient temperature for insulation layer construction should be ≥5℃, and enterprise standards requiring that the comprehensive cost of external wall construction should be controlled within 1000-1200 yuan / m². Both types of clauses will be accompanied by version and scope metadata.

[0056] To quantify the degree of deviation from soft constraints, a soft constraint violation function is defined. Let the actual value of a certain evaluation parameter be... The threshold is The upper and lower bounds of the interval are respectively and ,and When a soft constraint is satisfied, the degree of violation is specified. When not satisfied, Take the non-negative deviation amount from the threshold or interval boundary.

[0057] For the violation function of soft constraints Different calculation methods are used depending on the type of constraint, but they are uniformly expressed as the following piecewise function:

[0058] (1) When the soft constraint is a parameter, it should not be lower than the threshold. ,Right now hour:

[0059] ;

[0060] That is, when the actual parameter value is lower than the threshold, the difference is used. As a degree of violation; when the standard is met .

[0061] (2) When the soft constraint is a parameter, it should not be higher than the threshold. ,Right now hour:

[0062] ;

[0063] That is, when the actual parameter value exceeds the threshold, the difference is used. As a degree of violation; when not exceeding the limit .

[0064] (3) When the soft constraint is a parameter, it should be within the interval Internal time:

[0065] ;

[0066] That is, when the actual parameter value is lower than the lower bound At that time, the deviation value As a violation degree; when it is higher than the upper bound. At that time, the deviation value As a degree of violation; when within the interval .

[0067] Step S2: Construct an executable design syntax for wall design, establish a bidirectional mapping relationship between the executable design syntax and the building information model, and generate multiple candidate wall schemes that conform to the executable design syntax under the constraints of the constraint expressions.

[0068] This step constructs the executable design syntax WEDS-DSL for walls and a structural gene library, defining the layer sequence, material moisture resistance characteristics, airtightness rating, thermal bridge parameters, joint construction, and anchor density, and establishing a reversible mapping with Building Information Modeling (BIM) data. The executable design syntax WEDS-DSL includes formal definitions of wall layer sequence, material moisture resistance characteristics, airtightness rating, thermal bridge parameters, joint construction, and anchor density.

[0069] The interlayer thermal bridge parameters specify the type of thermal bridge, such as plate-end thermal bridges and beam-column thermal bridges, and the corresponding linear heat loss coefficients. Value, unit is W / ( The nodal line heat loss coefficient is defined separately for nodal areas such as window opening sides and wall corners, and is used in conjunction with inter-layer thermal bridge parameters for accurate thermal performance calculations. The equivalent thickness for moisture resistance is calculated using the following formula, which combines material moisture resistance and thickness into a unified equivalent thickness index, facilitating the assessment of moisture accumulation:

[0070] ;

[0071] In the formula, The equivalent thickness for the material's wet resistance; The material's wet resistivity is expressed in units of 1000 ppm. s / m; The thickness is measured in millimeters (mm).

[0072] Air tightness is classified into multiple levels according to standards, and is also associated with the air leakage rate per unit area, with units of [missing information]. Anchor density is defined as the number of anchors per unit area, with units of [number] anchors per [area]. m ², and at the same time specify the material grade of the anchor, such as stainless steel A2 grade.

[0073] The construction gene library is classified into material genes, construction node genes, and anchor genes.

[0074] Material genes are classified according to functional layers: the base layer includes 4 types, with a dry density range of 500-1200. kg / m ³, single-area, single-material; the insulation layer includes 5 types, with a thermal conductivity range of 0.024-0.040. The moisture barrier is designed to be matched to the climate zone; it includes four types with a thickness ranging from 1.5 to 4.0 mm and is only placed between the base layer and the insulation layer; the finishing layer includes five types with a density ranging from 1200 to 2800. kg / m ³, with corresponding auxiliary materials.

[0075] The structural node genes include: window opening node, insulation extension ≥200mm, Value ≤ 0.08W / ( At wall corners, insulation overlap should be ≥150mm, and thermal performance should match that of adjacent walls; at the junction of floor slabs and exterior walls, insulation should extend ≥300mm downwards, and expansion joints should be provided.

[0076] Anchor types include: expansion bolts, made of 304 / 316 stainless steel, 8-12mm in diameter, with a density of ≥4 pieces / unit when the height is >50m. m ²; Adhesive-anchored combination, using epoxy resin adhesive, bonding area ≥ 500mm², single service area ≤ 0.5. .

[0077] The construction gene bank follows a logical combination of material layering → node adaptation → anchor fixing: First, select materials in the order of base layer → moisture-proof layer → insulation layer → finishing layer. The base layer is a single type of material with matching strength in a single area. The thickness of the insulation layer is determined according to the climate zone. The moisture-proof layer is sandwiched between the base layer and the insulation layer. Then, use the same insulation layer material for window opening extension, wall corner overlap, and floor slab and exterior wall junction treatment, and match the corresponding node auxiliary materials. Finally, select expansion bolt type or adhesive anchor type anchors and fix them according to the height or bonding requirements.

[0078] The bidirectional mapping relationship between the executable design syntax and the building information model is achieved through preset mapping rules. During mapping, a one-to-one correspondence is first established between the design parameters defined in the executable design syntax and the attributes of the BIM model. Table 1 shows a partial correspondence between the fields of the executable design syntax and the fields of the BIM model.

[0079] Table 1. Partial Correspondence between Executable Design Syntax Fields and BIM Model Fields

[0080]

[0081] During compilation, executable design syntax parameters are automatically written into the corresponding attribute fields of the BIM model. During decompilation, attribute data is extracted from the BIM model and converted into parameters that conform to the executable design syntax. Data verification logic is set during the mapping process. When the executable design syntax parameters are inconsistent with the attribute data of the BIM model, such as the insulation layer thickness in the executable design syntax deviating from the corresponding layer thickness in the BIM model by more than 5%, a prompt message is automatically generated and manual review is triggered to ensure data consistency.

[0082] The method of generating multiple wall candidate schemes that conform to the executable design syntax using a large language model includes the following steps: converting constraint expressions into structured instructions through prompt engineering, calling material genes, node genes and anchor genes in the construction gene library; based on the retrieved gene data, using the large language model to combine and generate candidate schemes that conform to the executable design syntax specification, and outputting clause-construction-performance mapping information.

[0083] The selected large language model is based on a general large model, but needs to be fine-tuned for the field of building wall design. The fine-tuning data covers building energy conservation code provisions, wall construction cases, executable design syntax rules and clause-construction-performance mapping samples to ensure that the model accurately understands the domain terminology and design logic and avoids the domain bias of the general model.

[0084] The specific implementation logic for generating an executable design syntax scheme under constraint expressions using a large language model is as follows:

[0085] S1: Constraint Expression Pre-parsing. This process transforms constraint expressions into structured instructions understandable by the large language model. Constraint expressions include hard constraints, soft constraints, and penalty functions. Hard and soft constraints are broken down into three categories: "must-satisfy parameter thresholds + priority optimization objectives + violation penalty rules." A constraint list module with embedded prompt words clarifies the insurmountable boundaries and optimization directions of the large language model's generated scheme.

[0086] S2: Enhanced Real-Time Parameter Matching. The large language model calls the gene library retrieval algorithm through a pre-defined API interface to accurately match parameters according to constraint expressions: For heat transfer coefficient constraints, it calls the material gene retrieval interface, inputs the climate zone and insulation layer conditions, and filters materials that meet the thermal conductivity requirements. Then, it uses a thermal calculation model to back-calculate the required insulation layer thickness to ensure that the heat transfer coefficient meets the standard. For anchor bearing capacity constraints, it calls the anchor gene retrieval interface, inputs the building height, and filters anchor types that meet the tensile bearing capacity and layout density requirements. The retrieval results are returned to the large language model in the format of gene ID + key parameters, serving as the parameter basis for scheme generation and avoiding the generation of invalid structures that exceed the gene library's scope.

[0087] S3: Structured generation of executable design syntax. The large language model logically combines the retrieved parameters based on the preset syntax template in the prompt words. The syntax template includes preset writing formats for layers, nodes, and anchors: first, the matching material parameters are written into the layer sequence field in the order of base layer → moisture-proof layer → insulation layer → finishing layer; then, the node construction parameters and anchor configuration parameters are filled into the corresponding fields; finally, the target values ​​corresponding to each constraint are marked in the performance target field to ensure that the syntax conforms to the executable design syntax specification and that all parameters are within the constraint range.

[0088] S4: Real-time Constraint Compliance Verification. The large language model calls its built-in constraint verification algorithm module to check the parameters of the generated preliminary solution. This constraint verification algorithm module is implemented through the rule engine. As an independent algorithm module, the rule engine is pre-installed in the post-processing stage of the large language model's solution generation. Its core function is to transform the formal rules of hard and soft constraints in the constraint expression into directly executable parameter verification logic, replacing manual verification and ensuring that the solution parameters fully conform to the constraint requirements.

[0089] The specific implementation process of the rules engine and its adaptation to the patented technology logic are as follows:

[0090] (1) Rule preloading: Associating constraint expressions with verification logic. When the rule engine starts, it automatically reads the constraint expressions generated in step S1, which include hard constraints, soft constraints, violation functions, and penalty coefficients, and establishes a mapping relationship according to constraint type, constraint object, verification logic, and processing strategy to form a structured verification rule library. Hard constraint rules clearly define the constraint object, such as U-value, combustion performance rating, and threshold range, such as U-value ≤ 0.30. W / ( The system includes verification logic, checking whether parameters are within the threshold, and processing strategies. If a violation occurs, a re-search is triggered. Soft constraint rules clearly define the constraint objects, such as airtightness level, target value, such as ≥6 level, violation degree function, such as a quadratic penalty formula, and processing strategies. For minor violations, parameters are fine-tuned to reduce the penalty value.

[0091] (2) Scheme Parameter Extraction: Structured Analysis of Preliminary Scheme. The preliminary scheme generated by the large language model follows the executable design syntax (WEDS-DSL). The rule engine automatically extracts the core parameters in the scheme through the preset syntax parsing interface, forming structured data of parameter name-parameter value-associated gene ID, such as "insulation layer thickness - 80mm - material gene ID: M-003", "anchor density - 3 / - Anchor gene ID: A-001”, ensuring the accuracy of the verification object.

[0092] (3) Item-by-item verification and logical judgment. The rule engine verifies each extracted parameter in the order of hard constraints first, followed by soft constraints. During hard constraint verification, if the parameter meets the threshold requirement, such as the U value corresponding to an insulation layer thickness of 80mm = 0.28W / If the value is ≤0.30, it is marked as "compliant"; if it is violated, such as an anchor density of 3 / <4 / If the building height is greater than 50m, it will be marked as "serious violation" and a handling strategy will be triggered.

[0093] During soft constraint verification, if the parameters meet the standards, such as airtightness level 6, the violation degree is 0, and it is marked as "compliant"; if there is a minor violation, such as airtightness level 5, a penalty value is calculated using the violation degree function, such as... Marking it as a "minor violation" triggers a fine-tuning strategy.

[0094] (4) Verification result processing: closed-loop optimization scheme.

[0095] In the event of a serious violation (hard constraint violation), the rule engine returns a "parameter violation warning" and specific violation items to the large language model, such as "anchor density does not meet hard constraint: ≥4 anchors / anchors are required when height > 50m". This triggers the large language model to reconstruct the gene library and retrieve anchor genes that meet the density requirements, such as anchor gene ID: A-002, density = 4 / 4. Regenerate the scheme.

[0096] When minor violations (soft constraint violations) occur, the rule engine returns "parameter fine-tuning suggestions" to the large language model. For example, "Air tightness level 5, it is recommended to adjust the air tightness layer material to gene ID: M-008, the air tightness level is expected to be improved to level 6, and the penalty value is reduced to 0". The large language model fine-tunes the parameters based on the suggestions without having to search the gene database again.

[0097] When fully compliant, the rules engine outputs "verification passed," and the solution directly proceeds to the next step of multiphysics performance preliminary evaluation.

[0098] If the parameters are found to be inconsistent with the hard constraint threshold, the algorithm will automatically return to the gene pool to search for better materials and adjust the parameters to meet the requirements. If a slight violation of the soft constraint is found, the violation will be reduced by fine-tuning the parameters, while a penalty term will be calculated to ensure that the penalty value is minimized. Only after the verification is passed will the final executable candidate solution be output, realizing the algorithmic closed loop of "constraint → search → generation → verification".

[0099] Step S3: Use the surrogate model to perform multiphysics performance evaluation on the candidate wall schemes, trigger the high-precision solver to verify the results based on the uncertainty evaluation, and select the set of compliant schemes.

[0100] This step employs a multiphysics surrogate model with uncertainty estimation to rapidly evaluate the thermal performance, thermal bridging, condensation and moisture accumulation, and carbon emissions at the material stage for candidate wall solutions. The surrogate model is selected from sparse Gaussian process models, physics-guided neural network models, or gradient boosting tree models. The specific selection logic and applicable scenarios are as follows:

[0101]

[0102] The model selection process when implementing this invention follows the following priority judgment logic: first, determine the core model based on the physical characteristics of the evaluation indicators and the amount of data.

[0103] Core performance indicators (U-value, Value, condensation risk): Sparse Gaussian process model is preferred to meet the requirements of uncertainty quantification;

[0104] Thermal-humidity coupling and transient performance indicators (moisture accumulation): Physically guided neural network models should be preferred to ensure physical consistency;

[0105] Auxiliary performance indicators (carbon emissions, cost) + preliminary evaluation of large-scale solutions: prioritize gradient boosting tree model to balance efficiency and accuracy.

[0106] This invention can employ a multi-model collaborative hybrid strategy: for example, a gradient boosting tree model can be used to quickly filter out obviously non-compliant solutions, such as those exceeding carbon emission limits or cost limits, and then a sparse Gaussian process model can be used to evaluate the core performance indicators (U-value, etc.) of the remaining compliant solutions. The system performs precise evaluation and uncertainty calculation on the values, and finally calls a high-precision solver to verify the schemes that meet the triggering conditions, ensuring both efficiency and reliability of the evaluation of core indicators.

[0107] The rules for setting the model switching threshold are as follows: when the amount of training data is less than 1000 sets, the gradient boosting tree model is disabled and the sparse Gaussian process model is selected first; when the evaluation index involves explicit physical laws, such as thermal-humid coupling and heat transfer formulas, the constraint logic of the physical-guided neural network model is forcibly embedded; when the evaluation time of a single scheme is less than 10ms, the gradient boosting tree model is selected first, and the uncertainty quantification advantage of the sparse Gaussian process model is abandoned and it is only used for non-core indicators.

[0108] During the model training phase, a training set was constructed using historical wall design data, thermal-humidity coupling simulation data, and actual measured data of physical walls. Model parameters were optimized through 5-fold cross-validation to ensure the model's prediction accuracy. The surrogate model outputs the mean and variance of the target index, or the mean and 95th quantile interval, where the quantile interval is calculated based on the probability distribution of the model's prediction results.

[0109] The target indicators include wall heat transfer coefficient, condensation risk index, moisture accumulation index, carbon emissions at the material stage, and linear heat loss coefficient.

[0110] Wall heat transfer coefficient The following calculations were performed using the series connection of the heat transfer resistance of the inner and outer surfaces with the thermal resistance of each layer of material:

[0111] ;

[0112] In the formula, The internal surface heat transfer resistance; The external surface heat transfer resistance; For the first i The thickness of the layer material; For the first i Thermal conductivity of the layer material.

[0113] Condensation risk is assessed through the internal surface temperature coefficient. Conduct an assessment:

[0114] ;

[0115] In the formula, The temperature of the inner surface of the wall; Outdoor air temperature; Indoor air temperature.

[0116] When the coefficient reaches the specified threshold, it is determined that there is no risk of condensation.

[0117] The cumulative humidity index is obtained by continuously collecting humidity data through wall humidity sensors and calculating the duration or cumulative change of humidity exceeding the critical value within a specific period.

[0118] Carbon emissions at the material stage are calculated by querying the carbon footprint database of building materials to obtain the carbon emissions per unit mass of each material and then summing them up with the amount of material used.

[0119] linear heat loss coefficient The surface heat flux density of the thermal bridge node is measured by a heat flux meter, and the temperature of the node and surrounding walls is measured by a thermocouple. The result is calculated according to the standard formula. .

[0120] The high-precision solver can be conditionally triggered for verification, including the following triggering conditions:

[0121] (1) The variance of the performance index output by the surrogate model is greater than or equal to the preset variance threshold. ;

[0122] (2) The confidence interval width output by the surrogate model is greater than or equal to the preset width threshold. ;

[0123] (3) The performance parameters of the candidate wall scheme are within the preset range of the hard constraint boundary.

[0124] in The value ranges from 0.01 to 0.05, corresponding to the variance thresholds of different target indicators, such as the U value. Set to 0.02 ; The value range is 0.03-0.10, corresponding to the confidence interval width threshold for different target indicators, such as... Value Set to 0.06 The high-precision solver verification is also triggered when the distance between the candidate solution parameters and the hard constraint boundary is ≤ ε. =5% × hard constraint threshold.

[0125] The high-precision solver includes a steady-state / transient thermal-humidity coupling solver, an ISO condensation verification solver, and a two-dimensional / three-dimensional thermal bridge numerical solver. During the verification process, a finer computational mesh and physical model are used to ensure the accuracy of the evaluation results and to perform constructability checks.

[0126] Step S4: Iteratively optimize the set of compliant solutions under the penalty mechanisms of hard and soft constraints, output the Pareto optimal solution set, parse the Pareto optimal solution set based on the executable design syntax, and generate design documents and compliance audit reports.

[0127] This step employs a hybrid optimization strategy to iteratively optimize the set of compliant solutions. This strategy combines hierarchical evolutionary algorithms, Bayesian optimization algorithms, and policy gradient algorithms. The hierarchical evolutionary algorithm is used for global topology search, the Bayesian optimization algorithm is used for local optimization of continuous parameters, and the policy gradient algorithm is used to dynamically adjust hyperparameters during the optimization process.

[0128] like Figure 2 As shown, a combination of hierarchical evolutionary algorithm, Bayesian optimization algorithm, and policy gradient algorithm is implemented through a three-order collaborative approach of global exploration, local optimization, and policy iteration:

[0129] First, a hierarchical evolutionary algorithm guides the global exploration, satisfying the optimization objective according to hard constraints, such as... Value and combustion performance → core performance, such as annual energy consumption and condensation risk → secondary objectives, such as LCC cost and material carbon emission stratification. Each stratum corresponds to a different evolutionary population. Through selection, crossover, and mutation operations, an initial set of solutions that meet hard constraints is generated, while invalid solutions that violate the infinite penalty rule are filtered out, providing a compliance basis for subsequent optimization.

[0130] Next, the Bayesian optimization algorithm constructs a Gaussian process surrogate model of the objective function based on the performance feedback of the preceding population, quickly locates the local region with the best performance, and efficiently screens potential optimal solutions by collecting functions such as EI expectation improvement for parameters such as moisture accumulation and material carbon emissions, thereby reducing the cost of high-precision calculations.

[0131] Finally, the policy gradient algorithm intervenes with a reinforcement learning framework, transforming the optimization process into agent-environment interaction. It uses the real-time optimization benefits of each objective, such as the reduction in energy consumption and the cost saving rate, as reward signals to dynamically adjust the population iteration direction of the hierarchical evolutionary algorithm and the surrogate model update frequency of the Bayesian optimization algorithm.

[0132] State space of reinforcement learning Defined as a combination of a construction parameter vector, a performance index vector, and a constraint violation vector:

[0133] ;

[0134] ;

[0135] ;

[0136] ;

[0137] ;

[0138] In the formula, To construct the parameter vector; The number of key construction parameters; A vector of performance metrics; To optimize the number of targets; To constrain the degree vector of violation; To constrain the quantity; This is an indicator function; it takes the value 1 if the constraint is violated, and 0 otherwise. This is a hard constraint threshold; It is used to quantify the degree of violation of all constraints and is the core vector parameter representing constraint compliance in the state space.

[0139] Action space Defined as the single fine-tuning magnitude of each construction parameter:

[0140] ;

[0141] ;

[0142] In the formula, For the first i The single fine-tuning magnitude of each construction parameter, and ; , These are the minimum and maximum values ​​that the parameter can take. It is used to quantify the magnitude of a single adjustment made by an intelligent agent to the core structural parameters of a wall design, and is a core quantitative indicator of the action space.

[0143] reward function Taking into account the weighted standardized benefits of each objective, as well as the penalties for hard and soft constraints:

[0144] ;

[0145] ;

[0146] In the formula, For the first k Each target weight; For standardized returns; The first action after the action k Project value; A fixed penalty coefficient is applied to the hard constraint; It is a set with hard constraints; The penalty coefficient can be updated for soft constraints; It is a set of soft constraints.

[0147] Both target weights and soft constraint penalty coefficients can be updated online based on optimization feedback. Target weights are updated based on the deviation between the actual and average returns of each target, directing optimization resources towards targets with greater potential for increased returns. Soft constraint penalty coefficients are updated based on the deviation between the actual violation frequency and the target violation frequency, achieving adaptive control of the penalty intensity.

[0148] Target weight The update formula is:

[0149] ;

[0150] ;

[0151] In the formula, This represents the current iteration step number; The learning rate; For the first Step 1 The actual returns of the project target; For average returns; It is a local minimum.

[0152] The update formula for the soft constraint penalty coefficient is:

[0153] ;

[0154] In the formula, The learning rate is the penalty coefficient. For the first Average frequency of soft constraint violations in the first 10 rounds; The frequency of the target violation.

[0155] The meta-learning mechanism extracts meta-knowledge from historical project data, generates initial target weights and penalty coefficients, and updates the target weights and soft constraint penalty coefficients online based on real-time performance feedback during the iterative optimization process.

[0156] The specific logic of meta-learning is as follows: First, the meta-parameters of historical similar items are retrieved as initial values ​​using a case-based reasoning (CBR) method. The similarity between historical items and the current item is calculated, primarily through multi-dimensional feature weighted matching. Let the feature vector of the current item be... The first in the historical project database The feature vectors of each project are The feature dimension is The similarity calculation process is as follows:

[0157] S1: Project characteristics are categorized into three main types. Constructed in several dimensions:

[0158] ;

[0159] In the formula, Environmental characteristics, including climate zone, building height, and building type; The target characteristics include energy consumption targets, carbon emission limits, and cost budgets; These are constraint features, including the number of hard constraints and the priority of soft constraints;

[0160] S2: Use the Analytic Hierarchy Process (AHP) to determine the weights of each feature. ,satisfy Recommended weight configuration:

[0161] ;

[0162] S3: Current Projects and Historical Projects The weighted Euclidean distance between them is:

[0163] ;

[0164] S4: Convert distance to similarity score :

[0165] ;

[0166] when At that time, the project was judged. It is highly similar to the current project.

[0167] S5: Filter by similarity ranking Extract the meta-parameters (target weights) from each historical project. Penalty coefficient ), weighted average based on similarity percentage:

[0168] ;

[0169] ;

[0170] In the formula, and These are the initial target weight vector and the initial penalty coefficient vector for the current project, respectively.

[0171] S6: Employ Model-Independent Meta-Learning (MAML) to rapidly fine-tune the initial meta-parameters in 1-2 rounds, adapting them to the specific scenario differences of the current project.

[0172] ;

[0173] In the formula, For the set of meta-parameters; To fine-tune the learning rate; For a small sample of the current project The loss function on.

[0174] Meta-learning first extracts meta-knowledge from historical wall design optimization cases, such as the weighting patterns of annual energy consumption and material carbon emissions under different climate zones, and the experience of adjusting penalty coefficients for different soft constraints. This historical optimization logic is then transformed into initial target weight templates and penalty coefficient benchmarks. In the real-time optimization phase, meta-learning continuously monitors the optimization feedback of the current solution: if a solution meets hard constraints but exceeds the condensation risk limit, it retrieves meta-knowledge from historical cases on "how to adjust weights when condensation risk is high," increasing the weight proportion of the condensation risk target while simultaneously increasing the penalty coefficient for the corresponding soft constraint; if a solution shows significant optimization effect on material carbon emissions but annual energy consumption is close to the threshold, it combines historical optimization experience in carbon emission-energy consumption balance to fine-tune the weights of both, avoiding excessive emphasis on carbon emissions that could lead to excessive energy consumption.

[0175] The adaptive optimizer obtains robustness evaluation through Monte Carlo sampling of annual meteorological disturbances, material parameter fluctuations, and construction errors, with a sampling size N≥1000. For annual meteorological disturbance sampling, solar irradiance and outdoor dry-bulb temperature are sampled using a triangular distribution within a ±10% fluctuation range; for material parameter fluctuation sampling, thermal conductivity is sampled using a uniform distribution within a ±15% fluctuation range and thickness within a ±5mm fluctuation range; for construction error sampling, airtightness leakage rate is sampled using a truncated normal distribution within a ±20% fluctuation range. Robustness evaluation is achieved by calculating a robustness score, which represents the percentage of sampled schemes that satisfy all hard constraints and whose key indicators are within the target range. A higher robustness score indicates stronger disturbance resistance. Finally, Pareto schemes are ranked from highest to lowest robustness score, with schemes having a robustness score ≥0.9 being given priority recommendation.

[0176] In the Pareto frontier, summaries of the construction sequence, key parameters, and compliance verification results are given for three representative solutions: robust, economic, and extremely energy-efficient.

[0177] Based on the executable design syntax parsing of the Pareto optimal solution set, the system automatically generates wall construction specifications, key details of node drawings, and a bill of materials, forming a design document.

[0178] The compliance audit report is archived in both machine-readable and document formats. It includes a complete reasoning chain from the source of regulatory provisions, constraint expression transformation, performance evaluation data to the final compliance conclusion, along with version metadata. The report is archived in machine-readable format, such as JSON. The machine-readable file uses a structured data organization method, storing clause information, calculation basis, evaluation results, and optimization conclusions according to preset fields for automatic computer reading, retrieval, and analysis. A corresponding PDF visual report is also generated for manual review and archiving. The content of both formats is consistent to ensure cross-verification during the audit.

[0179] During the operation phase, measured data of the building are collected, and an adaptive window detection algorithm is used to handle discrete data drift, while a distribution difference statistical test algorithm is used to handle continuous data drift. When drift or performance degradation is detected, the model is updated and re-optimization suggestions are generated.

[0180] For continuous IoT data, such as temperature, humidity, and energy consumption data of the inner and outer surfaces of walls, the KS test method is used. First, probability distributions are constructed for the collected real-time data and the initial baseline data, where the initial baseline data is either simulated data from the design phase or stable data from the initial stage of operation. Then, the KS statistics for the two distributions are calculated. When the statistic is ≥0.05, it is determined to be data distribution drift.

[0181] For discrete IoT data, such as the frequency of condensation and the number of equipment failures, an adaptive window monitoring (ADWIN) method is adopted. This method monitors data changes in real time by setting a sliding window, continuously comparing the distribution differences between old and new data within the window. When the difference exceeds a preset threshold, it is determined to be data distribution drift. When distribution drift or performance degradation exceeds the threshold, incremental updates of the agent model and knowledge graph are triggered. The performance degradation exceeding the threshold refers to the deviation between the actual performance indicators of the wall and the design target exceeding the allowable range.

[0182] Triggering model updates includes: filtering incremental data after drift occurs, constructing a hybrid training set to fine-tune the proxy model; identifying conflicts between new data and existing knowledge graphs, updating entity attributes or inference rules in the knowledge graph based on data timeliness priority and domain specification constraints, and generating difference annotations.

[0183] For fine-tuning the proxy model, firstly, select IoT incremental data that meets the quality requirements after drift detection, and construct a mixed training set with the historical core dataset at a ratio of 7:3. Retain the original model network structure and only update the model parameters. For layers related to continuous performance indicators, use a small learning rate (1e-5~1e-4) for iterative training, and freeze the fixed layers related to hard constraint logic to avoid failure of core constraint judgment. Add a performance verification step during training. If the model prediction error exceeds the preset threshold after fine-tuning, backtrack to the data screening step to remove outliers until the model accuracy meets the standard.

[0184] For fine-tuning the knowledge graph, the incremental data is first broken down into entities, relationships, and attributes: entities include new material performance degradation data and equipment operating status, and relationships correspond to performance indicators and environmental factors; then, through the knowledge fusion module, the new entities / relationships are matched with the existing nodes of the graph, and for conflicting information, the data timestamp priority is combined with domain specifications for verification to determine whether to retain or correct it; finally, the graph reasoning rules are updated and difference annotations are generated to ensure that the graph is synchronized with the actual operating status.

[0185] During incremental updates, only the new data samples after the drift are used to fine-tune the model without retraining the entire model. At the same time, difference labels are generated, which include the time of drift, data type, degree of drift, specific indicators of performance degradation, and changes in model parameters before and after the update. The difference labels are archived along with the update records, making it easy to trace the cause of drift and the effect of the update.

[0186] Example

[0187] Taking the design of the exterior wall of a newly built office building in a cold region as an example, this case study selects a 15-story office building in the severe cold zone B, at 45° north latitude, with a winter outdoor design temperature of -26°C. The exterior wall design must simultaneously meet the requirements of energy conservation, safety, and economy.

[0188] Configuration parameters: =0.02, =0.06, ε =5% × threshold N =1000.

[0189] Objective: To meet mandatory requirements and minimize heating energy consumption.

[0190] Comparison of design phase time: The target environment and constraint compilation phase took 0.5 hours, achieved through automatic platform data collection and constraint compilation, eliminating the need for manual retrieval and interpretation; constraints were directly converted into functions callable by the optimizer. The executable design syntax generation phase took 0.8 hours, achieved through large language model generation and automatic BIM mapping; the construction combination was intelligently generated by the model, avoiding experience bias. The two-level evaluation phase took 3.5 hours, achieved through rapid initial evaluation using a proxy model and on-demand high-precision solution; the proxy model replaced 70% of the high-precision simulation, with only boundary solutions requiring review. The compliance audit report generation phase was extremely short, achieved through automatic generation and dual-format archiving; the entire audit chain was traceable, eliminating the need for manual processing. The total time was approximately 4.8 hours. The operation log of this method is shown in Table 2, and the intermediate data of this method and the traditional method are shown in Table 3.

[0191] Table 2 Running Log

[0192]

[0193] Table 3 Intermediate Data

[0194]

[0195] Results Comparison: Compared to the pure high-precision solution which took 14.6 hours and had a pass rate of 82.1%, this method took approximately 4.8 hours, had an average absolute percentage error of approximately 2.3%, and a pass rate of approximately 91.7%. Compared to the solution that only substitutes without verification, the accuracy and pass rate are significantly improved. Robustness is approximately 0.94.

[0196] Efficiency Improvement: The traditional method takes 14.6 hours to complete the design of one project, while this method only takes 4.8 hours. The time spent in the specification retrieval → constraint conversion step is reduced by 90%, and the time spent in the performance evaluation step is reduced by 60%. The core reason is that the collaborative replacement of manual operation by machine-readable constraints and proxy models.

[0197] Improved pass rate: The traditional method has an overall pass rate of 82.1% for 50 candidate solutions, and requires manual correction of hard constraint violations. This method has a pass rate of 91.7%, requiring only minor adjustments to soft constraint biases. This is mainly due to: ① Constraint compilers eliminating text interpretation biases; ② Hard constraint logic is embedded when generating solutions using a large language model; ③ Two-level evaluation filters out boundary risk solutions in advance.

[0198] Traceability: Traditional compliance verification requires compiling a 15-page manual report, including simulation screenshots and excerpts of regulatory clauses. This method automatically generates a JSON+PDF dual-format report, which can directly trace the entire chain from regulatory clauses to constraint expressions, surrogate model evaluation data, and conclusions, reducing audit time from 2 hours to 0.1 hours.

[0199] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0200] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A smart design method for building walls based on a semantic-physical dual closed loop, characterized in that, Includes the following steps: Step S1: Obtain project data and environmental element data of the target project, generate semantic tags and project context, retrieve specification clauses based on semantic tags, and compile the specification clauses into constraint expressions containing hard constraints, soft constraints and violation degree functions; The constraint compiler compiles the specification clauses into constraint expressions using a constraint compiler, which includes a clause vector indexing unit, a clause parser, and a constraint compilation unit. The text vector indexing unit performs text vectorization processing on the texts in regulations, atlases and enterprise standards and establishes an index library; The text parser extracts the target text from the index and decomposes it into a variable set, applicable scope, and performance threshold requirements. The constraint compilation unit constructs a dictionary for the building wall domain, performs dependency parsing on natural language texts, and extracts constraint objects, constraint relationships, thresholds, and applicable scope. The constraint compilation unit transforms the hard constraints into a combination of first-order predicate logic and mathematical formulas and associates them with an infinite penalty function through a logical rule mapping algorithm, and transforms the soft constraints into formal expressions and associates them with a quadratic penalty function. Step S2: Construct an executable design syntax for wall design, establish a bidirectional mapping relationship between the executable design syntax and the building information model, and generate multiple candidate wall schemes that conform to the executable design syntax under the constraints of the constraint expressions; Step S3: Use the surrogate model to perform multiphysics performance evaluation on the candidate wall schemes, trigger the high-precision solver to verify the results based on the uncertainty evaluation, and select the set of compliant schemes; Step S4: Iteratively optimize the set of compliant solutions under the penalty mechanisms of hard and soft constraints, output the Pareto optimal solution set, parse the Pareto optimal solution set based on the executable design syntax, and generate design documents and compliance audit reports; The set of compliant solutions is iteratively optimized using a hybrid optimization strategy, which includes a hierarchical evolutionary algorithm, a Bayesian optimization algorithm, and a policy gradient algorithm. The hierarchical evolutionary algorithm is used for global topology search, the Bayesian optimization algorithm is used for local optimization of continuous parameters, and the policy gradient algorithm is used for dynamically adjusting hyperparameters during the optimization process. The target weights and penalty coefficients are updated online through a meta-learning mechanism. This mechanism extracts meta-knowledge from historical project data and generates initial target weights and penalty coefficients. During the iterative optimization process, the target weights and the penalty coefficients of the soft constraints are updated online based on real-time performance feedback.

2. The intelligent design method for building walls based on semantic-physical dual closed loops according to claim 1, characterized in that: The executable design syntax described in step S2 includes formal definitions of wall layer sequence, material moisture resistance characteristics, air tightness level, thermal bridge parameters, node construction, and anchor density.

3. The intelligent design method for building walls based on semantic-physical dual closed loops according to claim 1, characterized in that: In step S2, the candidate wall schemes are generated. The process includes the following steps: converting the constraint expression into structured instructions through a prompting engineering approach, and calling up material genes, node genes, and anchor genes from the construction gene library; Using a large language model, candidate wall schemes that conform to the executable design syntax are generated based on the retrieved gene data combination, and the clause-construction-performance mapping information is output.

4. The intelligent design method for building walls based on semantic-physical dual closed loops according to claim 1, characterized in that: The conditions that trigger the high-precision solver verification in step S3 include: (1) The variance of the performance index output by the proxy model is greater than or equal to the preset variance threshold; (2) The confidence interval width output by the proxy model is greater than or equal to a preset width threshold; (3) The performance parameters of the candidate wall scheme are within the preset range of the hard constraint boundary.

5. The intelligent design method for building walls based on semantic-physical dual closed loops according to claim 1, characterized in that: The compliance audit report mentioned in step S4 is archived in both machine-readable and document formats. The compliance audit report includes a complete reasoning chain from the source of the regulatory provisions, constraint expression transformation, performance evaluation data to the final compliance conclusion, as well as version metadata.

6. The intelligent design method for building walls based on semantic-physical dual closed loops according to claim 1, characterized in that: Step S4 also includes a model update and optimization step: during the operation phase, measured data of the building is collected, an adaptive window detection algorithm is used to handle discrete data drift, and a distribution difference statistical test algorithm is used to handle continuous data drift. When data drift or performance degradation is detected, incremental data after the drift occurs is selected to construct a hybrid training set to fine-tune the proxy model, identify conflicts between new data and existing knowledge graphs, update entity attributes or inference rules in the knowledge graph based on data timeliness priority and domain specification constraints, and generate difference annotations and re-optimization suggestions.

Images