Intelligent Design Methods For Architectural Shapes And Structural Topology Based On Machine Learning

Architectural design is highly intellectual and creative work.It is so interesting and fascinating to explore artificial intelligence(AI)-based architectural design methodology,i.e.exploring the creativity of machines from a design perspective.Intelligent design methods of architecture are an important way to promote the whole engineering design industry from labor-intensive to technology-intensive in the future.However,the current architectural shape design relies heavily on experience and has no quantitative representation of architectural shapes,and the structural topology design involves complex and large scales,all of which hinder the intelligent design of architecture towards real-time interaction and aesthetically efficient intelligent design.Therefore,it is necessary to innovate the existing architectural and the structural topology design methods to achieve a more intelligent and more scientific architectural shape design and structural design.This will break the current boundaries between architects and structural engineers and pave the way for an intelligent design of buildings with the integration of aesthetics and mechanics in the future.The mathematical principles and operations of autonomous architectural shape design and automatic structural topology design are the core scientific problems that need to be studied to achieve the intelligent design.To this end,this paper systematically investigates intelligent design methods for architectural shapes and structural topology.The main contents involved are as follows:The Self-Sparse Generative Adversarial Network(Self-Sparse GAN)for autonomous generation of architectural shapes and the intelligent design method of architectural shapes based on Self-Sparse GAN without constraints are proposed,which overcomes the difficulties of no public datasets,no explicit geometry function,and no reasonable evaluation metric for architectural art shapes.Numerical experimental results show that the proposed intelligent design framework of architectural shapes learns the probability distribution of architectural shape features and can generate tens of thousands of sketches for architects in a few seconds.The generated sketches can not only learn the corresponding textures of the building but also place the textures in the correct positions.A novel AI-based design method of architectural shapes with user preferences is proposed,where this method can control the sketch generation in two aspects: shape preferences and architectural fusion.To achieve these two functions,the AI-based design method contains the shape preference module with Self-Sparse GAN and the architecture fusion module,thus allowing users to control the generation of architectural sketches through shape preference and architectural fusion.The shape preference module allows input of hand-drawn sketches or text,and proposes a selfsimilarity score(SS-score)to evaluate the degree of similarity of shapes.The architectural fusion module allows for the implementation of user-preference elements(such as artistic style and local culture)with the generated sketches,and proposes the line loss with a feature-based attention mechanism to enhance the clarity of the fusion sketch.The design results show that the proposed AI-designer can not only generate many desired architectural sketches with a specific theme while remaining diverse in just a few seconds but also form clearer sketches with architectural fusion by re-rendering using user-specified elements.A topology description function enhanced neural network for efficient explicit structural topology design(TDF-NN)is proposed,where cleverly combining neural networks and topology description functions.By introducing the Sigmoid activation function,the TDF values are converted into continuous density values of the material distribution,making the gradient of the objective function with respect to the TDF easy to calculate and allowing the use of gradient-based optimization algorithms to improve computational efficiency.The results of several 2D and 3D examples show that the proposed TDF-NN can achieve a speedup factor of approximately 2 to 10 compared to conventional topology optimization methods.A multi-resolution topology description function enhanced neural network for structural topology design(MR-TDF-NN)is proposed,which comprises three parts:multiresolution model to establish the relationship between the Knot mesh,Density mesh,and Displacement mesh;physical model solver to solve the state field of topology optimization.and model optimization to search the optimal parameters of MR-TDF-NN by optimizing the loss function.The proposed MR-TDF-NN can perform finite element analysis on a coarser finite element mesh,which reduces the computational effort of finite element analysis and further improves the computational efficiency of TDF-NN.The experimental results demonstrate that the MR-TDF-NN method is a versatile and efficient intelligent design method for structural topology.Compared to traditional multi-resolution topology optimization methods,it achieves speedup factors of approximately 10 to 18 times.In comparison to the TDF-NN method,it achieves speedup factors of approximately 11 times.Furthermore,it can achieve speedup factors of 20 to 100 times when compared to the traditional SIMP method.An integrative intelligent design framework for architectural shapes and structural topology is proposed(termed as Civil Engineering Intelligent Designer),which contains four modules: the intelligent design module of architectural shapes,the geometry modeling module,the intelligent design module of structural topology,and the architecture intelligent rendering module.The geometry modeling module extracts the geometric contours of architectural shaps,automatically discretizes complex structural shapes,and generates finite element meshes.The architecture rendering module decomposes the image rendering process into multiple small probability modeling steps to improve the modeling capability for complex probability distribution pairs and generate architectural renderings.The proposed method can automatically give a large number of diverse architectural shape sketches that meet user requirements without relying on human experience and automatically perform structural topology design under the given architectural shape to quickly generate multiple structural design solutions.Then,the architecture intelligent rendering module can rapidly give the corresponding architectural renderings.The proposed framework for integrative intelligent design for architectural shapes and structural topology explores the integrative design of shapes and structural topology of architecture for the first time,realizes the whole chain of the intelligent design process of creative design,structural topology design,and architecture rendering,and provides a methodological and theoretical foundation for more intelligent architecture design in the future.