MACHINE LEARNING-BASED DECISION SUPPORT FRAMEWORK FOR BIM COMPONENT SPECIFICATION

Building Information Modeling (BIM) environments contain structured, data-rich models, yet the design logic embedded within them is rarely reused beyond individual projects. Although BIM systems capture relationships between spatial configurations and building components, this knowledge typically remains project-specific and is not systematically utilized in future design processes. This study introduces a machine learning (ML)-based computational framework that predicts door specification attributes from room-to-room spatial transitions, aiming to support data-informed decision-making in design. In this context, BIM is approached not only as a modeling environment but also as a source of transferable design knowledge. A dataset of 5,763 door instances was compiled from three Revit-based projects representing different building typologies: a hospital, an office, and a recreational facility. Spatial transitions were encoded using a Bag-of-Words (BoW) representation, and Random Forest algorithms were applied for classification and regression. Four scenarios were tested: (i) intra-project learning, (ii) cross-project learning, (iii) combined project learning, and (iv) domain-augmented retraining. To provide a benchmark, model performance was evaluated against simple baseline strategies, including majority-class prediction for classification and mean-value prediction for regression. Results showed strong performance for some targets within single projects, while others exhibited more moderate results. Performance decreased across projects due to inconsistent naming, class imbalance, and label mismatch, where the model encounters previously unseen categories. The findings demonstrate that archived BIM models can be transformed into predictive design intelligence, enhancing computational reasoning and efficiency within BIM-based workflows, while emphasizing the need for data standardization to achieve robust cross-project applicability.