A government-anchored graph learning framework for predicting embodied carbon of construction materials in data-scarce regions

초록

Embodied carbon assessment of construction materials is critical for achieving carbon neutrality, yet countries such as South Korea lack product-level life cycle inventory (LCI) databases despite having category-level government evaluation coefficients. This study proposes a government-anchored graph learning framework that combines cross-regional graph features with Korean government calibration coefficients to predict the carbon intensity of 39,372 Korean construction materials. A heterogeneous graph integrating 52,486 processes from five international LCI databases is constructed through cross-lingual semantic matching. A region-adaptive Graph Attention Network with gated feature bypass and domain adaptation transfers carbon knowledge from data-rich regions to the unlabeled Korean domain, while conformal prediction provides distribution-free prediction intervals with finite-sample marginal coverage guarantees. The model achieves R2 = 0.735 on the labeled test set (5-fold CV: 0.734 ± 0.027), and per-category calibration anchored to Korean Environmental Product Declaration (EPD) evaluation coefficients (KEITI) yields Spearman ρ = 0.878 (p < 0.001) against 18 Korean construction materials with a median absolute error of 24.5% and 66.7% of predictions within ±50% of government reference values. The calibration extends Korean government category-level coefficients to 39,372 product-level predictions. A counterfactual analysis separates the two roles: the KEITI anchors—not the choice of predictor—supply the absolute, region-specific scale and are the primary source of Korean-side accuracy, but government coefficients exist for only 11 of the 20 unified categories; the GNN supplies coverage and generalization, degrading gracefully on the 72.3% of Korean products that have no confident cross-database match, where retrieval-based alternatives collapse (R2 ≈ 0). A QR code-based feedback system with Bayesian corrections further improves EPD compliance from 66.7% to 83.3% with 33 field observations. The framework offers a practical complementary system for carbon-informed material selection in data-scarce markets where category-level government data exists but product-level assessments are unavailable. Copyright

키워드