Enriching design-stage digital building models with component-level carbon emission information: a machine learning approach for laminated slabs

Despite the potential of precast concrete buildings to reduce lifecycle carbon emissions, design-stage digital building models remain largely limited to geometric and basic material information, making key attributes including component-level production carbon emissions difficult to obtain directly. This limits timely low-carbon decision-making in early design. Focusing on laminated slabs, this study develops a machine learning-based method to infer component-level carbon emissions and integrate quantitative carbon information into design-stage digital building models. Three models, namely BPNN, RF, and XGBoost, were compared for predicting direct and indirect production-stage carbon emissions. Results show that XGBoost achieved the best performance, with R2 values of 0.818 and 0.759 for direct and indirect emissions, respectively. Lightweight techniques and dimensionality reduction further reduced computational complexity with minimal accuracy loss, with R2 decreasing by less than 0.7%. Micro-level analysis identified component standardization rate as a key design descriptor for carbon reduction. To facilitate practical implementation, a SketchUp-integrated plugin was developed to automatically extract component information, predict carbon emissions in real time, and map results back to corresponding model components. The proposed method enriches design-stage digital building models with inferred carbon information, enabling practitioners to quantitatively translate design choices into measurable carbon savings and support proactive low-carbon decision-making.