A parametric generative design method for architectural space

Abstract

The present invention discloses a parameterized generative design method for architectural space, which solves the drawbacks that the current architectural engineering scheme design relies on the design experience of engineers, which is easy to cause scheme defects and the degree of intelligence needs to be improved. The key points of the technical scheme are to generate a two-dimensional or three-dimensional space scene by spatial modeling through a parameterized modeling method, including assigning component spatial parameters; the spatial scene is input into an evaluation model through feature transformation, and the evaluation model gives an evaluation value or evaluation grade and generates a reward and penalty value required by a reinforcement learning model; then reinforcement learning is performed through the reward and penalty values ​​to achieve parameter generation and optimization, and the optimized parameters are re-modeled until the optimization target is reached to obtain the most effective space design scheme. The parameterized generative design method for architectural space of the present invention can realize intelligent evaluation and self-optimization of architectural space and architectural design, and can effectively promote the development of digitalization and intelligence in the field of civil engineering.

Description

Technical Field

[0001] The invention relates to civil engineering technology, and in particular to a parametric generative design method for architectural space. Background Art

[0002] With the rapid development of urban construction in my country, intelligent construction has become the future development direction of the construction industry, and the degree of digitalization and intelligence of construction projects still needs to be improved. Especially in the design stage, the proposal of the plan often relies on the design experience of engineers. Once the engineers lack available experience, it is bound to cause defects in the plan. Therefore, in order to improve the quality of above-ground buildings and underground spaces, enhance safety, comfort and other functions, and reduce the experience defects in the space design stage, it is urgent to effectively promote the development of digitalization and intelligence in the field of civil engineering. Summary of the invention

[0003] The purpose of the present invention is to provide a parametric generative design method for architectural space, which can realize intelligent evaluation and self-optimization of architectural space and architectural design, and can effectively promote the development of digitalization and intelligence in the field of civil engineering.

[0004] The above technical objectives of the present invention are achieved through the following technical solutions:

[0005] A parametric generative design method for architectural space includes the following steps:

[0006] S1. Establish an initial model through parametric method, select basic components and their corresponding parameters required for design, and generate a preset space scene;

[0007] S2, the spatial scene is evaluated by inputting the evaluation model through feature conversion, giving the evaluation value / evaluation level and generating the reward and penalty value;

[0008] S3, the spatial scene is input into the reinforcement learning model and features are extracted. The reinforcement learning model accepts the reward and penalty values ​​output by the evaluation model, performs reinforcement learning, generates and updates the parameters of each component of the spatial scene, and optimizes the parameters.

[0009] S4, re-performing parametric modeling based on the optimized parameters to generate the corresponding spatial scene;

[0010] S5. Repeat steps S2-S4 until the optimization goal is reached and the optimal space design solution that meets the requirements is obtained.

[0011] Preferably, step S1 specifically includes:

[0012] Assign component parameters and form a spatial distribution matrix based on component position parameters and preset space;

[0013] Match the component information in the standard component library, place standard components in each spatial distribution matrix grid, and transform and assemble them through the transformation module to form a spatial scene.

[0014] Preferably, the evaluation model in step S2 includes:

[0015] Evaluation indicators, formulate quantitative criteria, and form multi-level comprehensive evaluation indicators;

[0016] The evaluation method, based on the established evaluation index system, conducts quantitative scoring and weighting of each spatial scene item by item, obtains scores and grades for specific plans and designs, and generates the reward and penalty values ​​required for reinforcement learning.

[0017] Preferably, the reinforcement learning model includes a strategy network, an encoder network and an action value network;

[0018] The encoder is used to extract features of the spatial scene, and the extracted scene features are input into the value strategy network;

[0019] The strategy network outputs the mean and standard deviation of each parameter and the entropy value of the strategy function according to the extracted scene features, and randomly samples according to the Gaussian distribution, and then clips the sampling results to obtain the corresponding component parameters;

[0020] The parameters generated by the policy network are re-parameterized and converted into spatial scenes, input into the evaluation model, and fed back to obtain reward and penalty values;

[0021] The action value network receives the evaluation results of the evaluation model, integrates the reward and penalty values ​​and the entropy value of the policy function, adjusts the parameters of the encoder and policy network, and optimizes and updates the reinforcement learning model.

[0022] In summary, the present invention has the following beneficial effects:

[0023] By combining modeling software, parametric modeling is realized, which greatly reduces the time cost required for manual modeling and makes the modeling accuracy controllable and efficient; combined with the evaluation model, real-time evaluation of the generated model is realized and feedback is given to the reinforcement learning model. Based on the reinforcement learning model, the automatic generation of component parameters and self-optimization of strategies can be realized, thereby improving the generation efficiency of underground space design solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 To generate a flow chart;

[0025] Figure 2 This is a flowchart of the reinforcement learning model;

[0026] Figure 3 Develop a process flow chart for parameterized modeling;

[0027] Figure 4 Schematic diagram of the spatial distribution matrix of component parameters. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0029] According to one or more embodiments, a method for parameterized generative design of architectural space is disclosed, such as Figure 1 As shown, the following steps are included:

[0030] S1. Establish the initial model through parametric method, select the basic components and their corresponding parameters required for design, and generate the preset space scene.

[0031] Select the required basic components and the corresponding sizes to establish a preset three-dimensional or two-dimensional space, including assigning parameters such as component spatial distribution, structural materials, and color. Divide the three-dimensional space into N×N cubic spaces with a side length of L, corresponding to the N×N dimension of the building space distribution matrix. Or divide the two-dimensional plane into N×N planes, and generate the distribution matrix of the plane space accordingly. Place standard components in the corresponding positions of each spatial distribution matrix grid. The standard components are defined and generated by design parameters such as spatial coordinates, materials, and colors, thereby forming a three-dimensional or two-dimensional space and its scenes.

[0032] S2. The spatial scene is evaluated by inputting the evaluation model through feature conversion, giving an evaluation value / evaluation level and generating reward and penalty values.

[0033] S3. The spatial scene is input into the reinforcement learning model and feature extraction is performed. The reinforcement learning model accepts the reward and penalty values ​​output by the evaluation model, performs reinforcement learning, generates and updates the parameters of each component of the spatial scene, and performs parameter optimization.

[0034] The reinforcement learning model mainly consists of three important parts: strategy network, encoder, and action value network. Figure 2 As shown. The parameter update and optimization of the model mainly includes the following parts:

[0035] 1) The encoder is responsible for extracting features from the preset spatial scene and converting the extracted scene features S t Input to the policy network.

[0036] 2) The strategy network is based on the scene features S extracted by the encoder t , output the mean μ corresponding to each parameter t and standard deviation σ t , and randomly sample according to Gaussian distribution, and cut the sampling results to get the corresponding component parameters P t .

[0037] 3) The parameters generated by the policy network are converted into spatial scenarios through modeling software, such as Figure 3 According to the position parameter information of the component and the preset space, a spatial distribution matrix is ​​formed, as shown in Figure 4 As shown in the figure, the space is divided into N×N cubic spaces with a side length of L (if the design space is two-dimensional, it is divided into N×N planes), and corresponding components are placed in each spatial distribution matrix grid, such as the top plate, bottom plate, central column, environmental facilities, lighting system, side wall, etc. By combining component parameter information with the standard component library, a spatial scene model is assembled.

[0038] 4) Comprehensively evaluate the quality of the generated spatial scene, and feed the evaluation results back to the action value network of the reinforcement learning model to promote the updating and optimization of the entire model. The evaluation model consists of two parts: evaluation indicators and evaluation methods. The evaluation indicators focus on safety, comfort, convenience, sustainability, etc., and further quantitative criteria are formulated in each aspect, and finally a multi-level comprehensive evaluation indicator is formed. In terms of evaluation methods, through intelligent methods such as expert systems, the spatial scenes are quantitatively scored and weighted item by item based on the established indicator system, and finally the specific plans and designs are scored and graded.

[0039] 5) The newly generated architectural space scene is sent to the evaluation model, and the corresponding reward and penalty value r(s) is obtained through the evaluation model. t ,a t The action value network is responsible for receiving the results of the evaluation model and synthesizing the reward and penalty value r(s t ,a t ) and the performance of the policy function entropy value H are used to regulate the parameters of the encoder and policy network, so that the entire reinforcement learning model can be optimized and updated. As shown in formula (1), the temperature parameter α is used to indicate the importance of entropy relative to the reward and penalty values.

[0040]

[0041] In formula (2), r represents the reward score, score represents the evaluation or grade score, and r td is the reward and penalty threshold, and n is the imbalance coefficient.

[0042] r=(score-r td ) n (1)

[0043] S4. Re-parametric modeling is performed based on the optimized parameters to generate the corresponding spatial scene.

[0044] S5. Repeat steps S2-S4 until the optimization goal is reached and the optimal space design that meets the requirements is obtained. The reinforcement learning model generates parameters and accepts evaluation feedback, continuously updates the policy network until the generated results are stable and the cumulative reward value is high. By continuously generating parameters and continuously evaluating, the model is self-updated and finally converges. At this time, it is considered that an optimal space design that meets the requirements is obtained.

[0045] Based on the reinforcement learning model, the automatic generation of component parameters and self-optimization of strategies can be achieved, which improves the efficiency of generating underground space design solutions. By combining modeling software, parametric modeling can be achieved, which greatly reduces the time cost required for manual modeling and makes the modeling accuracy controllable and efficient. By combining the underground space quality evaluation standards and intelligent algorithms, real-time scoring and evaluation of the generated model can be achieved and fed back to the reinforcement learning model, providing quantitative standards for model optimization and ensuring that the optimized underground space design has better quality.

[0046] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A parametric generative design method for architectural space, characterized in that: The following steps are included: S1. Establish an initial model through parametric method, select basic components and their corresponding parameters required for design, and generate a preset space scene; S2, the spatial scene is evaluated by inputting the evaluation model through feature conversion, giving the evaluation value / evaluation level and generating the reward and penalty value; S3, the spatial scene is input into the reinforcement learning model and features are extracted. The reinforcement learning model accepts the reward and penalty values ​​output by the evaluation model, performs reinforcement learning, generates and updates the parameters of each component of the spatial scene, and optimizes the parameters. S4, re-parametric modeling is performed according to the optimized parameters to generate the corresponding spatial scene; S5, repeating steps S2-S4 until the optimization goal is reached and the optimal space design scheme that meets the requirements is obtained; Step S1 specifically includes: Assign component parameters and form a spatial distribution matrix based on component position parameters and preset space; Match the component information in the standard component library, place standard components in each spatial distribution matrix grid, and transform and assemble them into a spatial scene through the transformation module; The evaluation model in step S2 includes: Evaluation indicators, formulate quantitative criteria, and form multi-level comprehensive evaluation indicators; Evaluation method: Based on the established evaluation index system, the spatial scenes are quantitatively scored and weighted item by item, and the scores and grades of specific plans and designs are obtained, and the reward and penalty values ​​required for reinforcement learning are generated; The reinforcement learning model includes a strategy network, an encoder network and an action value network; The encoder is used to extract features of the spatial scene, and the extracted scene features are input into the value strategy network; The strategy network outputs the mean and standard deviation of each parameter and the entropy value of the strategy function according to the extracted scene features, and randomly samples according to the Gaussian distribution, and then clips the sampling results to obtain the corresponding component parameters; The parameters generated by the policy network are re-parameterized and converted into spatial scenes, input into the evaluation model, and fed back to obtain reward and penalty values; The action value network receives the evaluation results of the evaluation model, integrates the reward and penalty values ​​and the entropy value of the policy function, adjusts the parameters of the encoder and policy network, and optimizes and updates the reinforcement learning model.

Images