DfMA-oriented rebar design optimization for reinforced concrete structures using graph neural network

Abstract

Reinforced concrete (RC) structure is the most commonly used structure format in buildings and civil infrastructures, wherein steel reinforcement (also called “rebar”) is embedded into concrete to enhance the tensile strength of concrete. Rebar detailing design is a critical stage of RC building design in terms of structural integrity, serviceability, durability and construction cost. In industrial practice, structural engineers first conduct structural analysis for the building based on a preliminary structural layout, using structural analysis software. The sizes of RC components are then adjusted, while the required steel areas for these components are identified based on the structural analysis results for rebar design. Currently, the rebar design is carried out manually or semi-automatically with some customized programs, which is time-consuming and relies heavily on structural engineers’ expertise and experience to achieve the optimal solution. Some research efforts have been made to automate this process, but some major limitations exist that hinders their applications in real- life projects: (1) impractical formulation of rebar design optimization problem, (2) lack of considerations on buildability of rebar design, (3) low computational efficiency, (4) lack of applicability to different component types, and (5) lack of consideration on interference between RC components. Echoed with the above limitations in current practices and existing studies, this research aims to develop an automatic and practical rebar design optimization approach with high computational efficiency considering buildability and interference between RC components. Design for Manufacture and Assembly (DfMA) and graph neural network (GNN) are the two key techniques identified from literature review to realize the above objective. Contributions lie in the following four aspects. (1) Chapter 3: Developed a DfMA-oriented rebar design optimization approach considering both material efficiency and buildability. Firstly, the rebar DfMA principles are first summarized through a review of rebar-related activities. Then, the rebar design optimization problem is explicitly formulated with detailed definition of design variables, constraints and objective functions for elongated RC components (including beams and columns). A quantitative approach for evaluation of buildability is developed and incorporated into the objective function for multi- objective optimization. The implement details are then introduced, including the method for rebar layout searching and the Exploratory Genetic Algorithm (EGA) for robust optimal solution searching. (2) Chapter 4: Developed a GNN-based approach for rapid rebar design optimization. Firstly, graph representations for elongated RC components are developed according to the characteristics of their typical rebar layouts and design methods to enable the adoption of GNN. Then, the rapid rebar design proposal using GNN is developed, which could immediately provide a near-optimal rebar design for a given design case. Since the rebar design proposed by GNN may not satisfy all the code requirements and the optimality of it is also not guaranteed, a post-processing algorithm is designed to check and optimize the initial design from GNN. (3) Chapter 5: Extended the DfMA-oriented and GNN-based rebar design optimization approach for RC flat components. Firstly, the DfMA-oriented optimal design formulation is extended for flat components (including slabs and walls) considering strip-based design method and finite element analysis. Secondly, the graph representations for the rebar design of flat components are constructed, based on which the GNN-based rebar design proposal approach is presented. Considering the hierarchy of design variables in flat components, the GNN is trained to predicted the positions of strip segmentation lines and further the rebar layouts in each strip. Finally, the MA-based checking and optimization algorithm is adjusted for flat components accordingly. (4) Chapter 6: Developed a proactive rebar clash avoidance approach based on explainable GNN and Rebar2Vec. Firstly, the representations of rebar clashes are introduced, including a vector representation and a graph representation. Based on this, the diagnosis of rebar clashes based on GNN is developed, which aims to not only identify rebar clashes but also classify rebar clashes as solvable or unsolvable clashes. Next, different strategies are developed to resolve solvable and unsolvable clashes. A Rebar2Vec model, which is a recommendation system for rebar layouts inspired by Word2Vec, is introduced to replace the rebar design efficiently to transform an unsolvable clash to a solvable clash. On the other hand, rebar reshaping and repositioning are employed to resolve a solvable clash. All the developed approaches are validated through experimental studies with real design cases. The results show that: (1) By incorporating DfMA principles into the design optimization of rebar, the buildability and practicality of final design can be improved. (2)GNN is good at capturing the interrelationship between different rebar groups and the incorporation with it can significantly reduce the computational time required by metaheuristic algorithms. (3) The developed DfMA-approach with GNN are applicable to different types of components including elongated and flat components. (4) Explainable GNN can be leveraged to efficiently identify and classify rebar clash, and a Rebar2Vec model enables the rapid rebar design adjustment to resolve clashes.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Jack Chin Pang CHENG (Supervisor)