top of page

BIM-based Generative Design (AI) 

BIM focuses on parameterisation and the automatic creation of a diverse set of new building layouts.
Generative AI, including Graph Convolutional Neural Network (GCNN), evaluates the fitness of each layout (Singh et al., 2022).​ 

Generative AI will help overcome the design flexibility and economy-of-scale constraints in Offiste.

Below is the proposed Generative AI framework to optimise manufacturing productivity in Offsite, along with the related scientific research.

We will test and further develop Generative AI frameworks to optimise DFMA.
Generative AI FINAL.png
​The objective of the research shown below is to develop an automatic building layout design algorithm that facilitates the process from generative design to prefabrication for precast buildings. 

  • The first study leverages BIM to automatically develop new design alternatives utilising the trained GCNN for efficient optimisation. In this study, GCNN seeks to evaluate the precast building layout that optimises wind effects. 

​​

  • A second study shows that the building design alternatives are evaluated for their performance in terms of manufacturing productivity. Discrete event simulation (DES) is employed to assess manufacturing productivity, leveraging its ability to model production processes. Ultimately, the design with optimised productivity is identified as the final design (Hexu Liu, 1 Yuxuan Zhang, 2 Zhen Lei, 3 Hong Xian Li, SangHyeok Han).

 

  • Generative AI and GCNN can automatically iterate and evaluate design alternatives that satisfy constructability requirements, such as structural integrity, industrialised construction processes, and transportation and logistics, in line with DFMA guidelines. ​​

  • The chart on the right exemplifies the proposed spatial-geometric parameters in a precast building with a complex layout. It also presents the building’s layout hierarchically, showing how individual modular flats are arranged spatially and combined into wings, which in turn form the entire floor plan. 

​​

  • After a precast building is parameterised using the spatial-geometric features, the unstructured building features are converted into a structured graph representation with MATLAB. This involves automatically mapping the modules’ spaces into a set of vertices, followed by converting the hierarchy of building attributes and spatial-geometric features into a graph representation. This involves constructing a connectivity graph to depict the spatial organisation of a precast building (consisting of multiple types of modularised flats), followed by allocating the spatial-geometric features to precast components within each modularised flat.

Graph neural network_edited.jpg

  • The BIM-based graph representation provides the basis for a graph-constrained precast layout generator for pre-construction planning and generative layout modelling.​​

  • As shown in the chart on the right, the precast layout generator leverages spatial-geometric features and a connectivity graph to generate new layout plans automatically.

  • Randomised iterative designs are generated using a BIM generative modelling algorithm, constrained by the shape and geometry of precast components and design constraints.

  • The output consists of new layout plans, spatial-geometric features, and connectivity graphs, which are converted into the feature matrix (X) and connectivity matrix (A).

  • The two matrices are required for a graph convolutional neural network (GCNN)-based discriminator to assess the optimality of the generative layout model.

  • Designing a building layout plan requires considering various factors, such as cost-effective materials, sustainability, and climate adaptability, including the influence of wind pressure on a building’s form and orientation, which is significant. 

  • The GCNN seeks to evaluate the precast layout, which optimises the wind effect.​

CGNN EXAMPLE 2.jpg

  • In a second study, the optimised design is determined by evaluating its performance with respect to manufacturing productivity.

​

  • To predict the production performance, a simulation model was developed, which mimics the production line. SimphonyNET is used to build the simulation model.

​​

  • The Des-based evaluation model simulates the manufacturing process to forecast the production time for each design and provides statistics on the prediction performance.

​

  • Ultimately, a design with optimised productivity is identified as the final design.

manufacturing process.jpg
bottom of page