Automated IFC Generation and Machine Learning-Based λ-Correction for Embodied Carbon Estimation of Buildings

This study proposes a practical framework for estimating embodied carbon in buildings by integrating automated generation of Industry Foundation Classes (IFC) models with X-correction techniques. Six minimal building attributes—gross floor area, number of floors, floor height, structural type, year of completion, and building use—were used to automatically generate simplified IFC models. Preset values for material thickness, density, and surcharge rates were applied, and Environmental Product Declaration (EPD) data were linked to calculate baseline emissions. The baseline IFC results, however, accounted for only 5-20% of actual embodied carbon, confirming systematic underestimation. To address this limitation, three correction methods were evaluated: global scaling with an average X, cohort-based correction using K-nearest neighbors, and machine learning regression with Light Gradient Boosting Machine (LightGBM). A dataset of 304 buildings, including 260 for training and 44 for testing, was used for validation. Results showed that scaling and cohort approaches provided limited accuracy, while the machine learning model achieved the best performance (R2=0.803, MAPE=23.3%). These findings demonstrate that even with minimal inputs, reliable embodied carbon estimation is feasible for existing buildings lacking design documents. The proposed framework supports BIM-LCA integration and contributes to data-driven strategies for carbon-neutral building design and retrofitting.