Bayesian enhanced graph neural networks: Refining design spaces for hollow concrete components with optimum mechanical performance

Abstract

Lightweight, non-structural concrete partition walls with customized hollow sections enhance earthquake resilience, particularly by optimizing horizontal stiffness to resist lateral forces in seismic engineering. However, exhaustive search of the entire design space to optimize horizontal stiffness is time-consuming. Therefore, this research aims to develop a classifier to filter out less promising sections, reducing the search space and improving efficiency. Combining Graph Neural Networks (GNNs) and Bayesian Neural Networks (BNNs) offers computational- and time-efficient solutions for classification tasks applicable to the design of hollow building components. However, the impact of different BNN configurations within this hybrid remains underexplored. To address this, we proposed and compared eight hybrid models with different BNNs to identify the optimal hybrid model based on classification accuracy. Results show that the BNN featuring a Bayesian layer followed by two linear layers (BLL) is most effective, achieving around 90 % classification accuracy in both training and testing datasets. To assess search space reduction, we test the models on 2000 samples. The hybrid model featuring BLL in its BNN achieves the best performance, with a 26.85 % search space reduction in the search space. Compared to a traditional statistical model, which achieves a 17.7 % search space reduction, the optimal hybrid model demonstrates superior effectiveness. The present study focuses on single-objective optimization, specifically targeting the performance of horizontal stiffness in directions more sensitive to the configuration change of morphed honeycomb channels in the hollow component. Future work will expand to multi-objective optimization, concurrently considering other mechanical properties for a more comprehensive optimization.