Empirical Quantification and Prediction of Building Component Lifespans with Graph Neural Networks

Extending the service life of existing buildings and their components is essential for slowing resource loops in circular construction. While overall building lifespans are relatively well understood, the actual service lives of individual components remain underexplored due to limited empirical data. This study addresses this gap by analyzing and modeling building component lifespans using real-world damage investigation records and condition assessments from Swedish buildings. Descriptive statistics reveal a wide range of lifespans across six major component categories, highlighting materials linked to damage from poorly designed construction details as a critical factor. Additionally, most residential components have markedly shorter lifespans-about 25-30 years-compared to non-residential ones. To predict component service lifespans, we configure building envelope deterioration factors as a generic acyclic graph and employ Graph Neural Networks for node-level regression. Among the tested models, Graph Convolutional Network achieved the highest performance (R² = 0.83), outperforming GraphSAGE and Graph Attention Network due to its effective uniform aggregation across neighboring nodes. Our findings provide a high-granular, data-driven approach to estimating component lifespans, which can enhance predictive maintenance strategies and optimize the reuse potential of reclaimed materials in their next lifecycles.