An Interpretable Machine Learning Framework for Structural Material Intensity Assessment of Academic Buildings Based on IFC-Derived BIM Data

Technology, Khulna, Bangladesh iftakherahamed2002@gmail.com Consequently, evaluation of structural material intensity (SMI) is critical for designing buildings sustainably, minimizing their embodied carbon content, and facilitating early decision-making processes related to the structure. In this context, traditional quantification and benchmarking approaches for structural material intensity continue to depend extensively on laborious processes, isolated calculation steps, and non- transparency of prediction models. As such, this paper introduces an explainable BIM-machine learning framework for material intensity assessment in the built environment based on IFC-enabled BIM modeling and benchmark analysis globally. For testing purposes, an informative BIM representation of the KUET Science Building in Bangladesh was created and exported in Industry Foundation Classes (IFC) format. With the help of a Python-based IFC parser, 498 wall components and 11 slab systems were automatically extracted along with relevant geometric and material attributes. Overall, these amounts resulted in an estimated structural volume of 34,389.47 m³ and composite structural material intensity (CSMI) of 2,714.64 kg/m². To create an informed predictive benchmark environment, a global dataset of 910 observations was collected and preprocessed, leading to the acquisition of 550 credible cases with 15 features relating to the structure and materials. Subsequently, feature engineering was applied to generate interpretable metrics such as the concrete-to-steel ratio, brick ratio, and wood ratio. Then, an XGBoost regressor model was employed for predicting CSMI, obtaining a high value of R² = 0.9857. To ensure better transparency and interpretation from an engineer's point of view, SHapley Additive exPlanations (SHAP) analysis was introduced into the framework. The output of the explainability analysis indicated that all concrete-related variables have a positive impact on structural material intensity, while timber-related attributes were shown to be negatively influencing the predicted CSMI values. Finally, a sustainability intelligence quadrant was defined as a way to compare the KUET Scie