Investigating on thermal insulation in concrete partition wall with insulation-material-infilled Voronoi sections: Feature impact analysis and efficient assessment approach

Reducing energy consumption in building operations, occupying 27 % of total carbon emissions, is crucial for decarbonization. Concrete partition walls with insulation-filled hollow sections, offer high thermal resistance, reducing air conditioning needs. Voronoi pattern represent a space being subdivided into polygonal regions and can be incorporated as the cavity pattern of the hollow section. However, the existing studies on Voronoi sections focus on porosity and pore quantity, with limited research on other feature metrics. This research addresses this gap by quantifying additional feature metrics of Voronoi sections using Bayesian inference, determining their collective influence on insulation. Findings reveal that feature metrics like cavity distribution and standard deviation of cavity areas affect insulation. Additionally, to improve time-consuming simulations, we proposed a model using Graph Neural Networks (GNNs) for accurate and rapid thermal assessments of complex Voronoi sections. Our GNN model, which uses graph-structured data rather than design parameters as training inputs, outperforms another four widely used models, achieving over 85 % prediction accuracy. In this research, two novel approaches are introduced: a Bayesian-based statistical model to understand the relationship between multiple feature metrics and insulation, pinpointing the most promising search space affiliated with certain values ranges of various feature metrics, and a rapid and accurate GNN-based model that uses graph representations of Voronoi sections in building components, preserving detailed geometric information for insulation prediction. Both approaches collectively improve the design process for building components with material-filled Voronoi sections by narrowing the search space and enhancing evaluation efficiency, respectively. • Impactful feature metrics affecting insulation of material-filled Voronoi section. • Bayesian inference for identification of impactful feature metrics. • Machine learning models for material-filled Voronoi section thermal assessment.