Exploring the impact of architectural and structural design on embodied carbon prediction using tree-based models and SHAP

Accurate estimation of embodied carbon in residential buildings is crucial for minimizing the environmental impacts of the built environment, particularly during the early design phases. However, conventional assessment methods are often time-consuming, data-intensive, and less adaptable in the early design stages, highlighting the need for predictive approaches such as machine learning. In this study, machine learning algorithms were employed to predict the embodied carbon of residential buildings based on process-based life cycle assessment (PB-LCA) data, using internationally recognized emission factors. Seven different models—Random Forest, XGBoost, LightGBM, Support Vector Regression (SVR), K-Nearest Neighbors (KNN), Gradient Boosting, and Multi-layer Perceptron (MLP)—were trained. After hyperparameter tuning, the four best-performing models were selected for detailed analysis. Beyond predictive accuracy, model interpretability was evaluated using SHapley Additive exPlanations (SHAP) to quantify and visualize the contribution of various input features. Results revealed that Gradient Boosting and Random Forest models emphasized infrastructural attributes such as concrete volume and foundation steel weight, while the XGBoost model showed a stronger focus on macro-level features like total floor area. LightGBM exhibited a hybrid behavior between these two groups. Moreover, non-linear patterns were observed in the influence of certain features—such as concrete volume and steel weight—on embodied carbon, particularly as their values increased. Building plan type and structural system also consistently emerged as key determinants across all models. The findings highlight the potential of advanced machine learning models, combined with interpretability tools like SHAP, not only to enhance prediction accuracy but also to support informed decision-making for architectural and structural optimization. This approach is particularly valuable for reducing environmental impacts during the early stages of building design. Finally, the study recommends incorporating local datasets and accounting for broader factors such as construction processes and supply chains to further improve model reliability and applicability.