A circularity-embedded explainable machine learning framework for environmental performance assessment of industrialized construction projects

Abstract

Industrialized construction projects require decisions that connect embodied carbon, material use, operation, and end-of-life recovery, yet these outcomes are rarely observed together. Existing stage-specific prediction and scoring studies therefore provide limited support for auditable project diagnosis under fragmented evidence. This study proposes CE-GP-XAI, a circularity-embedded, validation-aware explainable machine-learning framework that connects CLF WBLCA, BDG2, and CODD through disclosed scenario matching. Embodied-carbon intensity, material intensity, operational energy, and visual circularity recovery potential are normalized into a Green Performance Score, while SHAP and component contributions support verification and engineering diagnosis. The strongest task-specific results were obtained for Q75 high-carbon screening (AUC = 0.767), the historical-meter-free BDG2 S2 operational baseline (R 2 = 0.859), and independent evaluation of the official CODD checkpoint (mAP50−95 = 0.797). Quantitative matching achieved complete normalized-use coverage, a median absolute log-area distance of 0.528, and a log-area standardized mean difference of −0.325. Entropy-derived weights preserved the baseline ranking (Spearman ρ = 0.977) with 83.7% grade agreement. Joint uncertainty propagation produced a mean 95% GPS interval width of 26.87 and a mean modal-grade probability of 0.693. CE-GP-XAI provides a reproducible open-data pathway for comparative environmental screening, engineering diagnosis, and targeted evidence collection when jointly observed life-cycle data are unavailable. The integrated results should be interpreted as scenario-matched assessments rather than measured full-life-cycle assessments of the same physical projects.