3D printing concrete external wall thermal performance collaborative reverse design method based on conditional diffusion model

Abstract

The application provides a kind of 3D printing concrete outer wall thermal performance collaborative reverse design method based on conditional diffusion model: obtain multimodal data set, wherein each group of data contains the geometric image of building outer wall, and the nominal compressive strength and thermal insulation efficiency corresponding to the geometric image;The nominal compressive strength and thermal insulation efficiency are used as a condition vector to train the conditional diffusion model, so that the conditional diffusion model learns the mapping from the condition vector to the geometric image;The target nominal compressive strength and target thermal insulation efficiency input by the user are used as a condition vector to input into the trained conditional diffusion model, to generate the corresponding low-resolution topological representation;The low-resolution topological representation is input into the super-resolution reconstruction network to obtain the high-resolution topological representation;According to the high-resolution topological representation, print instructions for the additive manufacturing equipment are generated.The application can directly generate wall topological configuration that meets the bearing capacity and thermal insulation requirements at the same time, and can be directly used for 3D printing entity construction.

Description

Technical Field

[0001] This invention relates to the field of building exterior wall design and additive manufacturing technology, specifically to a collaborative reverse design method for the thermal performance of 3D printed concrete exterior walls based on a conditional diffusion model. Background Technology

[0002] Building exterior walls, as enclosure components with a very high heat transfer rate, urgently need to achieve significant heat transfer suppression and thermal bridging while maintaining high structural load-bearing capacity. Currently, existing technologies for improving the thermal insulation performance of exterior walls mostly rely on physical modifications, such as increasing the thickness of the exterior wall insulation layer, applying cast-in-place built-in insulation systems, or incorporating phase change materials and recycled aggregates into concrete. However, these traditional methods often fail to fundamentally achieve a balance between load-bearing capacity and high thermal insulation performance, and simply increasing insulation materials significantly increases construction costs and carbon emissions.

[0003] In terms of wall configuration design algorithms and methods, existing technologies mostly employ traditional topology optimization or forward parametric scanning. When faced with multi-objective collaborative optimization of mechanical properties (such as nominal compressive strength) and thermal properties (such as insulation efficiency), traditional algorithms suffer from significant drawbacks such as extremely high computational costs, long iteration times, and a high tendency to get trapped in local optima. Existing design methods struggle to efficiently reverse engineer the required wall topology configuration based on the set multi-objective performance. In recent years, some research has attempted to apply diffusion models to reverse design in other engineering fields, such as parametric inversion of stealth structures or energy-absorbing metamaterials. However, the application scenarios of these technologies are fundamentally different from the field of thermal-coordinated design of building exterior walls, and cross-domain migration has not yet been achieved. Furthermore, due to computational resource limitations, the structural images generated by existing reverse models often have blurred boundaries and grayscale transition zones, making it difficult for subsequent 3D printing slicing software to accurately identify geometric contours and printing paths, thus preventing direct use for high-precision manufacturing of solid components. The interior of traditional walls is usually solid or a single-chamber configuration, lacking complex chambers and blocking structures that can effectively impede heat transfer. Even if a complex heat insulation frame is designed through optimized algorithms, it is still difficult to achieve integrated molding due to the limitations of traditional mold manufacturing processes.

[0004] In summary, existing technologies have the following shortcomings: First, building exterior wall insulation methods remain at the stage of physical material stacking, failing to innovate structurally at the internal topological framework level; second, advanced generative AI reverse engineering technology has not yet been applied to the multi-objective collaborative design of building exterior walls; third, the structural images generated by existing diffusion models have low resolution and blurred boundaries, making them unsuitable for direct path planning in additive manufacturing. Therefore, there is an urgent need in this field for a method that can directly reverse engineer wall topological configurations that simultaneously meet load-bearing capacity and thermal insulation performance requirements without adding physical insulation materials, and ensure that the generated configurations have clear physical boundaries to support the construction of 3D printed entities. Summary of the Invention

[0005] The purpose of this invention is to provide a synergistic reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model. This method aims to address the technical problems of existing exterior wall insulation technologies relying on physical material stacking, which makes it difficult to balance load-bearing capacity and thermal insulation performance; existing reverse generation technologies not being applied to multi-objective synergistic design of exterior walls; and the resulting structures having ambiguous boundaries that prevent direct use in 3D printing manufacturing. The specific technical solution is as follows: A collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model includes: Obtain a multimodal dataset, wherein each set of data in the multimodal dataset includes geometric topographic images of building exterior walls, and nominal compressive strength and thermal insulation efficiency corresponding to the geometric topographic images; The nominal compressive strength and thermal insulation efficiency are used as conditional vectors to train the conditional diffusion model, enabling the conditional diffusion model to learn the mapping from the conditional vectors to the geometric shape image. The target nominal compressive strength and target thermal insulation efficiency input by the user are used as conditional vectors and input into the trained conditional diffusion model to generate the corresponding low-resolution topological representation. Inputting low-resolution topology representations into a super-resolution reconstruction network yields high-resolution topology representations. Printing instructions are generated based on high-resolution topology characterization for execution by additive manufacturing equipment.

[0006] Preferably, obtaining a multimodal dataset specifically involves: Design a variety of different building exterior wall configurations, which include inner walls, outer walls, and internal components located between the inner and outer walls. By changing the wall thickness of the inner wall, the wall thickness of the outer wall, the shape of the internal components, the wall thickness of the internal components, and the position of the internal components, different geometric configurations can be generated. Mechanical and thermal simulations were performed on the generated geometric configurations to obtain the nominal compressive strength and thermal insulation efficiency. Collect geometric topography images of all geometric configurations, along with their corresponding nominal compressive strength and thermal insulation efficiency, to obtain a multimodal dataset.

[0007] Preferably, the nominal compressive strength and thermal insulation efficiency are used as conditional vectors to train the conditional diffusion model, specifically: Constructing the training sample dataset: Geometric images are resized and binarized to form an initial image matrix. , the initial image matrix Corresponding nominal compressive strength and insulation efficiency as a condition vector ; where the initial image matrix The black areas in the diagram correspond to concrete material areas, and the white areas correspond to blank areas. Training the conditional diffusion model: Randomly select the initial image matrix from the training sample dataset. and its corresponding condition vector Random sampling time step and Gaussian noise For the initial image matrix Forward diffusion yields a noisy image. Noisy images Time step and condition vector Input conditional denoising network outputs predicted noise To minimize prediction noise Compared with real Gaussian noise The error between them is used to update the network parameters for the target; in, For parameters Conditional denoising network, Indicates the first Noisy images at each time step , This is the preset total number of time steps for forward diffusion.

[0008] Preferably, when training the conditional diffusion model, peak signal-to-noise ratio (PSNR) and structural similarity are used as evaluation metrics to assess the topology reconstruction quality of the conditional diffusion model. Training is stopped when the changes in PNR and structural similarity within a consecutive preset number of rounds are both less than the corresponding preset changes.

[0009] Preferably, a corresponding low-resolution topological representation is generated, specifically: Based on the target nominal compressive strength input by the user and target insulation efficiency , forming condition vectors ; Sample the initial noise image from a standard Gaussian distribution. and in the condition vector Under the constraints, in accordance with Denoising is performed stepwise in reverse order up to 1 to generate a low-resolution topological representation. ; in, This is the preset total number of time steps for forward diffusion.

[0010] Preferably, the reverse denoising process is expressed as follows: in, , Represents the mean of an inverse distribution. Indicates the variance of the reverse process. Indicates the first The image obtained by reverse denoising at each time step. Indicates the first The image obtained by reverse denoising at each time step. Indicates a Gaussian normal distribution. Indicates based on the current noisy image and condition vector Predict the noisy image from the previous time step The probability distribution, This represents the parameters of the conditional denoising network.

[0011] Preferably, the inverse distribution mean Represented as: in, For the first Retention coefficients at each time step The cumulative retention coefficient from the first time step to the t-th time step. Represents the noise variance coefficient. This indicates that the conditional diffusion model is based on the conditional vector. Under the constraints of the first Prediction noise at each time step.

[0012] Preferably, when training the super-resolution reconstruction network, low-resolution topological representations and corresponding high-resolution topological representations are used to form paired samples, and the network parameters are updated by minimizing the reconstruction loss function. The reconstruction loss function is a weighted sum of absolute error loss and structural similarity loss.

[0013] Preferably, based on high-resolution topological characterization Generate printing instructions for the additive manufacturing equipment to execute, specifically: High-resolution topology representation Perform geometric mapping to obtain the basic unit of the horizontal section of the wall; The basic unit is copied horizontally to generate a continuous horizontal section of the wall. The horizontal cross-section of the wall is stretched in the height direction to obtain a three-dimensional configuration; Pre-processing of the three-dimensional configuration before manufacturing; The 3D configuration, after preprocessing before manufacturing, is sliced ​​and path planned to generate printing instructions.

[0014] Preferably, the three-dimensional configuration undergoes pre-processing before manufacturing, specifically: Smooth and de-alias the boundaries of the three-dimensional configuration; The wall thicknesses of the inner wall, outer wall, and internal components in the 3D configuration are compared with the width of the printing strip. If any one of them is not an integer multiple of the width of the printing strip, local geometric merging or adjustment is performed until the wall thicknesses of the inner wall, outer wall, and internal components are all integer multiples of the width of the printing strip.

[0015] The application of the technical solution of the present invention has the following beneficial effects: This invention uses nominal compressive strength and thermal insulation efficiency as conditional vectors to train a conditional diffusion model, enabling the model to directly and inversely generate building exterior wall topology configurations that meet target performance requirements from random noise. Unlike existing technologies that first determine the geometric configuration and then perform simulation iterations for optimization, this invention transforms multiphysics multi-objective optimization from iterative search to generative solutions under conditional constraints, significantly improving design efficiency, avoiding getting trapped in local optima, and enhancing the satisfaction of target performance.

[0016] In the reverse denoising process of the conditional diffusion model, this invention injects geometric topography, nominal compressive strength, and thermal insulation efficiency as multimodal conditional features into the conditional denoising network, subjecting the denoising process to both mechanical and thermal physical constraints. Compared to existing diffusion models in the construction field that only use images as conditions or process different physical fields separately, this invention achieves synergistic constraints on thermal performance, generating results that more closely resemble the actual physical targets, reducing biases caused by post-processing compensation, and improving the reliability and consistency of multi-physical performance collaborative design.

[0017] This invention, after generating a low-resolution topological representation using a conditional diffusion model, introduces a token-to-token super-resolution reconstruction network. Through a global contextual attention module, a residual token selection group, and a sub-pixel upsampling layer, it performs sub-pixel-level reconstruction of the topological boundaries, outputting a high-resolution binary topological image with clear physical boundaries. Compared to existing diffusion models that output low-resolution images with blurred boundaries and grayscale transition zones, this invention significantly improves boundary clarity and geometric recognizability, enhancing the accuracy of geometric contour recognition and path planning success rate in slicing software. This allows the algorithm's generated results to be truly implemented in printable physical components.

[0018] This invention converts the generated high-resolution topological representation into binary topological encoding, and then sequentially performs geometric mapping, pre-manufacturing preprocessing, slicing, and path planning to finally generate printing instructions that can be directly executed by additive manufacturing equipment. The entire process forms a complete end-to-end closed loop from performance target input to solid component output, realizing the integration of intelligent design and digital construction of building exterior walls, and breaking through the limitations of traditional mold manufacturing processes on the forming of complex topological configurations.

[0019] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the collaborative reverse design method for the thermal performance of 3D printed concrete exterior walls based on the conditional diffusion model in Example 1; Figure 2 This is a diagram of the architecture of the conditional diffusion model in Example 1; Figure 3 This is a diagram of the architecture of the super-resolution reconstruction network in Example 1; Figure 4 These are schematic diagrams of the finite element models of the four types of building exterior wall configurations in Example 1, where (a) is the basic polygonal finite element model, (b) is the longitudinal diaphragm polygonal finite element model, (c) is the multi-polyline finite element model, and (d) is the sinusoidal cavity finite element model. Figure 5 The AI ​​generation model is generated based on the condition vector using the design method of Example 1, where (a) is the basic polyline AI generation model, (b) is the longitudinal diaphragm polyline AI generation model, (c) is the multi-polyline AI generation model, and (d) is the sine wave cavity AI generation model. Figure 6 These are the nominal compressive strength-displacement response curves of the basic polyline finite element model and the basic polyline AI-generated model in Example 1; Figure 7 The three-dimensional stress distribution field of the basic polygonal building exterior wall configuration in Example 1 is shown in (a) and (b) is shown in the three-dimensional stress distribution field of the basic polygonal finite element model. Figure 8 These are the dynamic thermal response time curves of the inner and outer walls of the basic piecewise linear finite element model and the basic piecewise linear AI-generated model under periodic temperature boundaries in Example 1. Figure 9 It is the three-dimensional spatial temperature distribution field of the basic polygonal building exterior wall configuration in Example 1 during the low temperature stage, where (a) is the three-dimensional spatial temperature distribution field of the basic polygonal finite element model, and (b) is the three-dimensional spatial temperature distribution field of the basic polygonal AI generated model. Figure 10 These are the nominal compressive strength-displacement response curves of the longitudinal diaphragm polygonal finite element model and the longitudinal diaphragm polygonal AI-generated model in Example 1. Figure 11 It is the three-dimensional stress distribution field of the longitudinally diaphragm-shaped building exterior wall configuration in Example 1, where (a) is the three-dimensional stress distribution field of the longitudinally diaphragm-shaped finite element model, and (b) is the three-dimensional stress distribution field of the longitudinally diaphragm-shaped AI-generated model. Figure 12 These are the dynamic thermal response time curves of the inner and outer walls of the longitudinally diaphragm broken line finite element model and the longitudinally diaphragm broken line AI generated model under periodic temperature boundaries in Example 1. Figure 13 It is the three-dimensional spatial temperature distribution field of the longitudinally diaphragm-shaped building exterior wall configuration in Example 1 during the low temperature stage, where (a) is the three-dimensional spatial temperature distribution field of the longitudinally diaphragm-shaped finite element model, and (b) is the three-dimensional spatial temperature distribution field of the longitudinally diaphragm-shaped AI-generated model. Figure 14 These are the nominal compressive strength-displacement response curves of the multi-segmented finite element model and the multi-segmented AI-generated model in Example 1; Figure 15 The three-dimensional stress distribution field of the multi-segmented building exterior wall configuration in Example 1 is shown in (a) and (b) is shown in (b) and (c) respectively. Figure 16 These are the dynamic thermal response time curves of the inner and outer walls of the multi-segmented finite element model and the multi-segmented AI generated model under periodic temperature boundaries in Example 1. Figure 17 It is the three-dimensional spatial temperature distribution field of the multi-segmented building exterior wall configuration in Example 1 during the low temperature stage, where (a) is the three-dimensional spatial temperature distribution field of the multi-segmented finite element model, and (b) is the three-dimensional spatial temperature distribution field of the multi-segmented AI generated model. Figure 18 These are the nominal compressive strength-displacement response curves of the sinusoidal curved cavity finite element model and the sinusoidal curved cavity AI-generated model in Example 1; Figure 19 It is the three-dimensional stress distribution field of the sinusoidal curved cavity building exterior wall configuration in Example 1, where (a) is the three-dimensional stress distribution field of the sinusoidal curved cavity finite element model, and (b) is the three-dimensional stress distribution field of the sinusoidal curved cavity AI generated model; Figure 20 These are the dynamic thermal response time curves of the inner and outer walls of the sinusoidal curved cavity finite element model and the sinusoidal curved cavity AI generated model under periodic temperature boundaries in Example 1. Figure 21 It is the three-dimensional spatial temperature distribution field of the sinusoidal curved cavity building exterior wall configuration in Example 1 during the low temperature stage, where (a) is the three-dimensional spatial temperature distribution field of the sinusoidal curved cavity finite element model, and (b) is the three-dimensional spatial temperature distribution field of the sinusoidal curved cavity AI generated model. Detailed Implementation

[0021] To facilitate understanding of the present invention, a more comprehensive description is provided below, along with preferred embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the present invention.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0023] Example 1: See Figure 1 This implementation provides a collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model, including the following steps: Obtain a multimodal dataset, wherein each set of data in the multimodal dataset includes geometric topographic images of building exterior walls, and nominal compressive strength and thermal insulation efficiency corresponding to the geometric topographic images; The nominal compressive strength and thermal insulation efficiency are used as conditional vectors to train the conditional diffusion model, enabling the conditional diffusion model to learn the mapping from the conditional vectors to the geometric shape image. The target nominal compressive strength and target thermal insulation efficiency input by the user are used as conditional vectors and input into the trained conditional diffusion model to generate the corresponding low-resolution topological representation. Inputting low-resolution topology representations into a super-resolution reconstruction network yields high-resolution topology representations. Printing instructions are generated based on high-resolution topology characterization for execution by additive manufacturing equipment.

[0024] The design method of this embodiment will be described in detail below: Preferably, in this embodiment, obtaining the multimodal dataset specifically involves: First, various building exterior wall configurations are designed. These configurations include inner walls, outer walls, and internal components located between the inner and outer walls. The inner and outer walls serve as the main supporting framework, while the internal components divide the space between them into multiple passageways. Different geometric configurations are generated by changing the wall thickness of the inner and outer walls, the shape of the internal components, their wall thickness, and their position. Specifically, the internal components refer to the supporting structures located between the inner and outer walls, used to connect or separate spaces, such as diagonal braces, straight ribs, and curved walls.

[0025] In this implementation, the outer envelope dimensions of the building exterior wall configuration are set to a length of 40cm and a width of 20cm (this mainly considers the planar structural dimensions of the building exterior wall configuration; the three-dimensional structure of the entire wall can be obtained by stretching the planar structure in the height direction). By changing the wall thickness of the inner wall, the wall thickness of the outer wall, the shape of the internal components, the wall thickness of the internal components, and the position of the internal components of each building exterior wall configuration, 4000 geometric configurations are finally obtained.

[0026] Next, mechanical and thermal simulations were performed on the generated geometric configuration using finite element analysis software. In the mechanical simulation, an axial compressive load was applied to the geometric configuration to obtain the nominal compressive strength. : in, This refers to the load value that continues until the structure (i.e., its geometry) reaches its peak load capacity and exhibits significant damage or failure. The nominal interface area of ​​the wall is equal to the length of the geometric configuration multiplied by its width.

[0027] In thermal simulation, external dynamic temperature boundary conditions are applied to simulate the heat transfer process and obtain the insulation efficiency, which characterizes the thermal insulation performance of the wall. : in, This represents the highest temperature of the inner wall after the structure reaches its quasi-steady-state temperature. This represents the lowest temperature of the inner wall after the structure reaches its quasi-steady-state temperature. This represents the highest temperature of the outer wall after the structure reaches its quasi-steady-state temperature. This represents the lowest temperature of the outer wall after the structure reaches the quasi-steady-state temperature.

[0028] Finally, after completing mechanical and thermodynamic simulations for all geometric configurations, geometric morphology images and simulation data (i.e., nominal compressive strength) of all geometric configurations were collected. and insulation efficiency Each data point includes a geometric morphology image and nominal compressive strength. and insulation efficiency This forms a multimodal dataset.

[0029] Preferably, the nominal compressive strength and thermal insulation efficiency are used as conditional vectors to train the conditional diffusion model, specifically: Training samples for constructing the conditional diffusion model: The geometric images are uniformly sized and then binarized to form an initial image matrix. , the initial image matrix Corresponding nominal compressive strength and insulation efficiency as a condition vector ; where the initial image matrix The black areas in the image correspond to concrete material areas, and the white areas correspond to blank areas (i.e., areas without concrete material). In this embodiment, the geometric shape image is uniformly compressed to 32×64 pixels.

[0030] Establishing the forward diffusion process: using the initial image matrix Starting from, in Gaussian noise is gradually added within each time step to make the initial image matrix The image gradually transforms from a clear topology to random noise. Specifically, the forward diffusion process is represented as follows: in, Indicates the first Noisy images at each time step Indicates the first Noisy images at each time step Represents the noise variance coefficient. It is the identity matrix. , This is the preset total number of forward diffusion time steps. Indicates a Gaussian normal distribution. This indicates the noisy image from the previous time step. Get the current noisy image The probability distribution.

[0031] Furthermore, noisy images at any time step Represented as: in, Represents standard Gaussian noise, satisfying , Indicates a Gaussian normal distribution. It is the identity matrix; For the first Retention coefficients at each time step From the first time step to the second The cumulative retention factor at each time step. For the first The retention coefficient at each time step.

[0032] Constructing a conditional denoising network: denoising noisy images Time step and condition vector A common-input conditional denoising network is used to predict the noise added during the forward diffusion process. ; like Figure 2 As shown, the conditional denoising network employs a U-Net encoder-decoder structure. The network input is a noisy image of 32×64 pixels. The encoding path compresses the feature map from 32×64 pixels to 2×4 pixels through progressive downsampling, passing through feature layers of scales such as 16×32, 8×16, and 4×8 in sequence; the decoding path progressively upsampling restores the original resolution. Multi-scale features are transmitted between the encoder and decoder via skip connections; the downsampling operation in the encoding path preserves key information while compressing features. Each processing step receives a time step... The embedding of this allows the network to sense the current level of denoising during the diffusion process.

[0033] Specifically, the predicted noise output by the conditional denoising network is expressed as: , For parameters Conditional denoising network, nominal compressive strength and insulation efficiency These features can be converted into conditional features through fully connected layers or embedding layers and injected into the U-Net network along with time-step features, so that the model is subject to the joint constraints of mechanical and thermal performance during the denoising process.

[0034] Training the conditional diffusion model: Randomly select the initial image matrix from the training sample dataset. and its corresponding condition vector Random sampling time step and Gaussian noise For the initial image matrix Forward diffusion yields a noisy image. Noisy images Time step and condition vector Input conditional denoising network outputs predicted noise To minimize prediction noise Compared with real Gaussian noise The error between the two is used to update the network parameters.

[0035] Specifically, the training loss function in this embodiment is expressed as follows: in, Indicates prediction noise Compared with real Gaussian noise The error between them Represents the mathematical expectation. This represents the square of the L2 norm.

[0036] Furthermore, during the training of the conditional diffusion model, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) are used to evaluate the topology reconstruction quality of the conditional diffusion model, and the number of training iterations is determined based on the changes in the evaluation metrics. Specifically, after each training round or after several training rounds, the geometric topography image generated by the model is compared with the corresponding real geometric topography image in the dataset, and the PSNR and SSIM metrics are calculated.

[0037] The PSNR metric is expressed as follows: in, The maximum possible value of an image pixel. The mean square error between the generated geometric topography image and the real geometric topography image.

[0038] The SSIM metric is expressed as follows: in, Represents the generation of geometric topography images Images of real geometric shapes Structural similarity, To generate geometric topography images The mean, For true geometric shape images The mean, To generate geometric topography images variance For true geometric shape images variance To generate geometric topography images Images of real geometric shapes covariance, and This is a constant used to avoid the denominator being zero.

[0039] Specifically, PSNR is used to evaluate the pixel error between the generated geometric topography image and the real geometric topography image, while SSIM is used to evaluate the similarity between the two in terms of structural topography. As the number of training epochs increases, both PSNR and SSIM gradually increase and tend to stabilize, indicating that the conditional diffusion model can effectively learn the distribution characteristics of the building exterior geometric topography image and its relationship with the conditional vector. The mapping relationship between them. Furthermore, both PSNR and SSIM tend to stabilize, specifically: the changes in the peak signal-to-noise ratio and structural similarity index within a consecutive preset number of rounds (such as two consecutive rounds) are less than the corresponding preset changes.

[0040] In this embodiment, when the number of training epochs reached approximately 500, the PSNR stabilized at approximately 31 dB and the SSIM stabilized at approximately 0.97 or higher, indicating that the model training had essentially converged. Therefore, in this embodiment, the number of training epochs for the conditional diffusion model was set to 500; in other embodiments, the number of training epochs can be adjusted according to the convergence of PSNR and SSIM.

[0041] Preferably, after the conditional diffusion model is trained, it can be used to determine the target nominal compressive strength based on the user input. With target insulation efficiency Generate the corresponding low-resolution topological representation, specifically: Based on the target nominal compressive strength input by the user and target insulation efficiency , forming condition vectors ; Sample the initial noise image from a standard Gaussian distribution. and in the condition vector Under the constraints, in accordance with Denoising is performed stepwise in reverse order up to 1 to generate a low-resolution topological representation. ; Specifically, the reverse denoising process is represented as follows: in, , Represents the mean of an inverse distribution. Indicates the variance of the reverse process. Indicates the first The image obtained by reverse denoising at each time step. Indicates the first The image obtained by reverse denoising at each time step. Indicates a Gaussian normal distribution. Indicates based on the current noisy image and condition vector Predict the noisy image from the previous time step The probability distribution, This represents the parameters of the conditional denoising network.

[0042] Furthermore, the mean of the inverse distribution can be calculated from the noise or equivalent parameters predicted by the conditional diffusion model, denoted as: in, For the first Retention coefficients at each time step The cumulative retention coefficient from the first time step to the t-th time step. Represents the noise variance coefficient. This indicates that the conditional diffusion model is based on the conditional vector. Under the constraints of the first Prediction noise at each time step.

[0043] The final image obtained after reverse denoising This is a low-resolution topological representation. Specifically, it is a geometric topographic image with a size of 32×64 pixels. However, this low-resolution topological representation... The boundaries are too blurry to be directly used for high-precision 3D printing. Therefore, this embodiment introduces a Token-to-Token Super-Resolution Reconstruction Network (TTST) to address this issue. Enhance the clarity of boundaries.

[0044] like Figure 3 As shown, the super-resolution reconstruction network comprises, in sequence: an input convolutional layer, a global context attention module (GCA), multiple cascaded residual token selection groups (RTSGs), an output convolutional layer, and a pixel-shuffle layer. Each residual token selection group (RTSG) contains a parallel Top-k token selection attention branch (TTSA) and a window-based self-attention branch (WSA), and is coupled with an aggregation and multi-scale feedforward layer (MFL) and a global context attention (GCA).

[0045] Specifically, low-resolution topological representation (Size 32×64) First, the input convolutional layer maps the binary pixel values ​​to a high-dimensional feature map. This feature map is then fed into a Global Contextual Attention (GCA) module, which aggregates contextual information across the entire image, enabling features at each location to perceive the overall topology (e.g., the connectivity of chambers, the orientation of supporting skeletons), thus achieving shallow feature extraction and normalization. The normalized feature map then passes through multiple cascaded Residual Token Selection Groups (RTSGs). Within each RTSG, Top-k token selection attention and window-based self-attention are performed in parallel: the Top-k branch selects the k most important tokens from the feature map for sparse attention computation, focusing on key supporting skeletons and chamber boundaries in the topology; the window-based self-attention branch divides the feature map into non-overlapping windows, performing multi-head self-attention computation independently within each window to capture local thin-wall details and rounded corner transitions. The outputs of the two branches are fused and fed into an aggregation and multi-scale feedforward (MFL) layer. This layer extracts multi-scale local patterns through convolutional kernels or dilated convolutions of different scales, while further aggregating global information with global context attention. Multiple cascaded RTSGs progressively refine the boundaries and structural connectivity of features. The output of the last RTSG is processed by an output convolutional layer to adjust the number of channels to meet the requirements of pixel rearrangement. Finally, the pixel rearrangement layer performs sub-pixel-level upsampling and magnification operations, ultimately outputting a high-resolution image (128×256 pixels) magnified by 4 times, i.e., a high-resolution topological representation. .

[0046] Specifically, in this embodiment and All implementations employ binary topological coding, with black areas corresponding to concrete material regions and white areas corresponding to blank regions (i.e., regions without concrete material). In some embodiments, multi-channel coding may also be used. If multi-channel coding is used, the probability or category of different structural regions is encoded separately, and the output is reconstructed. Used for subsequent boundary vectorization and geometric fitting.

[0047] Furthermore, high-resolution topological representations are used when training the super-resolution reconstruction network. With low-resolution topology representation These constitute paired samples. Depend on Obtained according to a preset downsampling strategy (e.g., bilinear interpolation downsampling). Inputting the data into the TTST network, the reconstructed data is obtained through forward propagation. The network parameters are updated by minimizing the reconstruction loss function, so that... Approaching The reconstruction loss function is expressed as: in, and All are weighting coefficients. For absolute error loss, For the generated high-resolution topological representation Compared with the original high-resolution topological characterization Structural similarity index between them The training loss function for the super-resolution reconstruction network.

[0048] Through the above training, the super-resolution reconstruction network can learn the mapping relationship from fuzzy low-resolution topology to clear high-resolution topology.

[0049] After completing the training of the super-resolution reconstruction network, the low-resolution topological representation generated by the conditional diffusion model is used. By inputting a trained TTST network, the corresponding high-resolution topological representation can be output. Compared to low-resolution topological representation High-resolution topological characterization With clear physical boundaries and accurate geometric contours, it can improve the accuracy and stability of boundary extraction and subsequent geometric processing, and can be directly used for geometric modeling and path planning in subsequent 3D printing.

[0050] Preferably, based on high-resolution topological characterization Generate printing instructions for the additive manufacturing equipment to execute, specifically: High-resolution topology representation Perform geometric mapping to obtain the basic unit of the horizontal section of the wall; The basic unit is copied horizontally to generate a continuous horizontal section of the wall. The horizontal cross-section of the wall is stretched along the height direction to obtain a three-dimensional configuration consistent with the printed specimen. ,in, This indicates the column coordinates of the scanned image along the lateral movement direction of the printing gun. This indicates the row coordinates of the scanned image along the front-to-back direction of the printing table. This indicates the vertical movement direction of the printing gun, representing the number of layers stacked on the wall. The 3D configuration undergoes pre-processing before manufacturing, specifically: smoothing and de-aliasing the boundaries to reduce boundary burrs caused by the increased topology sampling resolution; simultaneously, comparing the wall thickness of the inner wall, the outer wall, and the internal components with the printing strip width to ensure that the wall thickness is an integer multiple of the printing strip width. If it is not an integer multiple, the local geometry of the structure is merged or fine-tuned; where the printing strip width refers to the width of the concrete strip extruded by the 3D printing equipment after deposition.

[0051] The 3D configuration, after preprocessing before manufacturing, is sliced ​​and path planned to generate printing instructions.

[0052] Specifically, slicing and path planning include: importing the 3D configuration into slicing software (such as Simplify3D, Cura, PrusaSlicer, or Slic3r); setting the layer height, line width, path spacing, and printing speed according to the nozzle parameters of the concrete or mortar extrusion printer; determining the printing direction and layer cutting plane to ensure that the layer outline is consistent with the coordinates of the 3D configuration; generating the motion trajectory of the nozzle or extruder head and determining the outer contour and necessary inner filling path for each layer; and setting the printing material ratio and curing conditions.

[0053] After the above process is completed, printing instructions are obtained that can be directly executed by additive manufacturing equipment.

[0054] Finally, the printing instructions are imported into the printing equipment, and the target exterior wall solid component is formed by layer-by-layer deposition, resulting in a three-dimensional solid model. During the printing process, the process consistency is maintained according to the preset material delivery and curing strategy to achieve high-fidelity reproduction of the model's morphology. After printing is completed, necessary curing is performed to obtain the final specimen.

[0055] Verification Case: Furthermore, to verify the accuracy and generalization ability of the conditional diffusion model in complex multiphysics reverse design, this embodiment selects typical samples of four types of building exterior wall configurations—basic polygonal, longitudinally diaphragm polygonal, multi-polygonal, and sinusoidal cavity—from the multimodal dataset and constructs finite element models (FEA models) for each. Figure 4 (a) to Figure 4 As shown in (b); the nominal compressive strength of each typical sample was further extracted ( ) and insulation efficiency ( As input conditions, the conditional diffusion model is driven to generate AI generative models that satisfy the corresponding conditional vectors, such as... Figure 5 (a) to Figure 5 As shown in (d), the finite element model and the AI-generated model are respectively imported into Comsol Multiphysics (i.e., numerical simulation software) for forward solving, and the mechanical and thermal response characteristics of the finite element model and the AI-generated model are compared.

[0056] In terms of mechanical properties, by Figure 6 , Figure 10 , Figure 14 and Figure 18 It can be seen that the nominal compressive strength-displacement response curves of the finite element model and the AI-generated model show a high degree of overlap throughout the overall loading process. Among the four sets of comparisons, the peak strength ( The relative error is extremely small, for example in multi-segmented lines ( Figure 14 In the model, the nominal compressive strength of the finite element model is 16.52 MPa, while the nominal compressive strength of the AI-generated model is 16.40 MPa (with an error of only 0.7%); the largest deviation occurs in the sinusoidal cavity type ( Figure 18 In the model, the nominal peak compressive strengths of the two models were 17.61 MPa and 18.27 MPa, respectively, with errors controlled within 4.0%. However, it should be noted that although the peak loads of the finite element model and the AI-generated model are similar, in some models, due to the influence of geometric fine-tuning, there are still some differences in the nonlinear evolution stage of the load-displacement curves. Figure 5 The two-dimensional cross-sectional topologies of (a), (b), (c), and (d) reveal that the AI-generated model is not a simple copy of the original data, but rather incorporates geometric variations in wall thickness, brace angles, or curvature. Despite these local topological differences, Figure 7 , Figure 11 , Figure 15 and Figure 19 The three-dimensional stress distribution field confirms that the AI-generated model effectively reproduces the macroscopic force transmission structure and stress concentration mode similar to the finite element model, indicating that the conditional diffusion model has learned the deep physical mapping law between geometric features and bearing capacity.

[0057] In terms of thermal properties, Figure 8 , Figure 12 , Figure 16 and Figure 20 The dynamic thermal response duration curves of the inner wall (inner surface of the model) and the outer wall (outer surface of the model) under a periodic temperature boundary are given, where This represents the temperature fluctuation value of the inner wall in the finite element model. This represents the fluctuation value of the inner wall temperature in the AI-generated model. This represents the temperature fluctuations of the outer wall in the finite element model and the AI-generated model. It can be seen that the green dashed line representing the inner wall temperature of the AI-generated model closely matches the blue solid line representing the inner wall temperature of the finite element model, indicating that their insulation effects against drastic external temperature fluctuations are essentially similar. Quantitative data further confirms this conclusion: in the four comparison groups, the insulation efficiency... The absolute errors did not exceed 0.01 (for example, Figure 8 Thermal insulation efficiency of finite element model Thermal insulation efficiency of AI-generated models ; Figure 12 Thermal insulation efficiency of finite element model Thermal insulation efficiency of AI-generated models Furthermore, through comparison Figure 9 , Figure 13 , Figure 17 and Figure 21 During the low-temperature phase (i.e., when the outer temperature is lower than the inner temperature), the three-dimensional temperature distribution field shows that, regardless of whether it is the short thermal bridge of the basic polygonal type or the thermal blocking mechanism constructed by longitudinal ribs inside the longitudinally partitioned polygonal type, the generated structure exhibits good topological equivalence to the finite element model in terms of the construction of the solid-phase heat transfer path. The heat flow conduction trajectory inside the generated structure effectively conforms to the set thermal insulation target.

[0058] Comparing the generated geometric information of the finite element model and the AI-generated model reveals a high degree of similarity in their overall morphology. Particularly in key geometric metrics such as the normalized area of ​​the structure and the length and area of ​​the heat transfer path, the two models exhibit strong consistency, ensuring similar load-bearing capacity and thermal insulation performance at the corresponding mechanical and thermal levels. Furthermore, with... Figure 13 and Figure 17 For example, in the comparison, although the two models have opposite topological orientations of heat transfer channels at certain locations, the nominal compressive strength is area-dependent, while the thermal conductivity is mainly constrained by the length and area of ​​the heat transfer path. Therefore, the sensitivity of the aforementioned geometric metrics to channel orientation is relatively low. Thus, channel reversal does not significantly disrupt the consistency of target performance. This result shows that the conditional diffusion model can not only reproduce the appearance morphology, but more importantly, it learns and maintains the core geometric laws that determine mechanical load-bearing capacity and heat transfer path.

[0059] In summary, through multi-dimensional morphological comparison and quantitative physical index verification, it is demonstrated that the conditional diffusion model constructed in this embodiment possesses high reverse design accuracy. This conditional diffusion model can not only accurately generate the corresponding macroscopic topology based on specified multi-objective conditions (mechanical strength and thermal insulation efficiency), but also ensure that the generated configuration conforms to real physical laws in terms of force and heat transfer mechanisms, thus providing a reliable computational paradigm for achieving efficient structure-function integrated on-demand design.

[0060] Example 2: This embodiment provides a collaborative reverse design system for the thermal performance of 3D printed concrete exterior walls based on a conditional diffusion model, including a memory and a processor. The memory stores a computer program, and when the processor runs the computer program, it executes the collaborative reverse design method for the thermal performance of 3D printed concrete exterior walls based on a conditional diffusion model in Embodiment 1.

[0061] Example 3: This embodiment provides a storage medium storing a computer program, which, when run, executes the collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model, as described in Embodiment 1.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model, characterized in that, include: Obtain a multimodal dataset, wherein each set of data in the multimodal dataset includes geometric topographic images of building exterior walls, and nominal compressive strength and thermal insulation efficiency corresponding to the geometric topographic images; The nominal compressive strength and thermal insulation efficiency are used as conditional vectors to train the conditional diffusion model, enabling the conditional diffusion model to learn the mapping from the conditional vectors to the geometric shape image. The target nominal compressive strength and target thermal insulation efficiency input by the user are used as conditional vectors and input into the trained conditional diffusion model to generate the corresponding low-resolution topological representation. Inputting low-resolution topology representations into a super-resolution reconstruction network yields high-resolution topology representations. Printing instructions are generated based on high-resolution topology characterization for execution by additive manufacturing equipment.

2. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 1, characterized in that, Obtaining a multimodal dataset specifically involves: Design a variety of different building exterior wall configurations, which include inner walls, outer walls, and internal components located between the inner and outer walls. By changing the wall thickness of the inner wall, the wall thickness of the outer wall, the shape of the internal components, the wall thickness of the internal components, and the position of the internal components, different geometric configurations can be generated. Mechanical and thermal simulations were performed on the generated geometric configurations to obtain the nominal compressive strength and thermal insulation efficiency. Collect geometric topography images of all geometric configurations, along with their corresponding nominal compressive strength and thermal insulation efficiency, to obtain a multimodal dataset.

3. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 1, characterized in that, The nominal compressive strength and thermal insulation efficiency are used as conditional vectors to train the conditional diffusion model, specifically: Constructing the training sample dataset: Geometric images are resized and binarized to form an initial image matrix. , the initial image matrix Corresponding nominal compressive strength and insulation efficiency as a condition vector ; where the initial image matrix The black areas in the diagram correspond to concrete material areas, and the white areas correspond to blank areas. Training the conditional diffusion model: Randomly select the initial image matrix from the training sample dataset. and its corresponding condition vector Random sampling time step and Gaussian noise For the initial image matrix Forward diffusion yields a noisy image. Noisy images Time step and condition vector Input conditional denoising network outputs predicted noise To minimize prediction noise Compared with real Gaussian noise The error between them is used to update the network parameters for the target; in, For parameters Conditional denoising network, Indicates the first Noisy images at each time step , This is the preset total number of time steps for forward diffusion.

4. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 3, characterized in that: When training the conditional diffusion model, peak signal-to-noise ratio (PSNR) and structural similarity are used as evaluation metrics to assess the topology reconstruction quality of the conditional diffusion model. Training is stopped when the changes in PNR and structural similarity within a preset number of consecutive rounds are both less than the corresponding preset changes.

5. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 1, characterized in that, Generate the corresponding low-resolution topological representation, specifically: Based on the target nominal compressive strength input by the user and target insulation efficiency , forming condition vectors ; Sample the initial noise image from a standard Gaussian distribution. and in the condition vector Under the constraints, in accordance with Denoising is performed stepwise in reverse order up to 1 to generate a low-resolution topological representation. ; in, This is the preset total number of time steps for forward diffusion.

6. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 5, characterized in that, The reverse denoising process is represented as: in, , Represents the mean of an inverse distribution. Indicates the variance of the reverse process. Indicates the first The image obtained by reverse denoising at each time step. Indicates the first The image obtained by reverse denoising at each time step. Indicates a Gaussian normal distribution. Indicates based on the current noisy image and condition vector Predict the noisy image from the previous time step The probability distribution, This represents the parameters of the conditional denoising network.

7. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 6, characterized in that, The inverse distribution mean Represented as: in, For the first Retention coefficients at each time step The cumulative retention coefficient from the first time step to the t-th time step. Represents the noise variance coefficient. This indicates that the conditional diffusion model is based on the conditional vector. Under the constraints of the first Prediction noise at each time step.

8. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 1, characterized in that: When training the super-resolution reconstruction network, low-resolution topological representations and corresponding high-resolution topological representations are used to form paired samples, and the network parameters are updated by minimizing the reconstruction loss function. The reconstruction loss function is a weighted sum of absolute error loss and structural similarity loss.

9. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 1, characterized in that, Based on high-resolution topological characterization Generate printing instructions for the additive manufacturing equipment to execute, specifically: High-resolution topology representation Perform geometric mapping to obtain the basic unit of the horizontal section of the wall; The basic unit is copied horizontally to generate a continuous horizontal section of the wall. The horizontal cross-section of the wall is stretched in the height direction to obtain a three-dimensional configuration; Pre-processing of the three-dimensional configuration before manufacturing; The 3D configuration, after preprocessing before manufacturing, is sliced ​​and path planned to generate printing instructions.

10. The collaborative reverse design method for the thermal performance of 3D-printed concrete exterior walls based on a conditional diffusion model according to claim 9, characterized in that, Pre-processing of the three-dimensional configuration before manufacturing specifically includes: Smooth and de-alias the boundaries of the three-dimensional configuration; The wall thicknesses of the inner wall, outer wall, and internal components in the 3D configuration are compared with the width of the printing strip. If any one of them is not an integer multiple of the width of the printing strip, local geometric merging or adjustment is performed until the wall thicknesses of the inner wall, outer wall, and internal components are all integer multiples of the width of the printing strip.

Images