Calculation of embodied GHG emissions in early building design stages using BIM and NLP-based semantic model healing

To reach the goals of limiting global warming, the embodied greenhouse gas (GHG) emissions of new buildings need to be quantified and optimized in the very early design stages, during which design decisions significantly influence the success of projects in achieving their performance goals. Semantically rich building information models (BIM) enable to perform an automated quantity take-off of the relevant elements for calculating a whole building life cycle assessment (LCA). However, imprecise type and property information often found in today's BIM practice hinders a seamless processing for downstream applications. At the same time, the early design stages are characterized by high uncertainty due to the lack of information and knowledge, making a holistic and consistent LCA for supporting design decisions and optimizing performance challenging. In assessing this often vague information, it is essential to consider different levels of element and material information for matching BIM to LCA data. For example, the structural properties of concrete are not yet defined in early design stages and should instead be considered as a range of material options due to different compressive strength classes. This paper presents a novel methodology for automatically matching the coarse information available in BIM models of the early design stages to the respective entries in LCA databases as a basis for a fully automated calculation process of the embodied GHG emissions of new buildings. This approach solves the existing gap in the automation process of manually enriching BIM models and adding information of LCA data and missing layers of vague models. In more detail, the proposed method is based on Natural Language Processing (NLP), using different strategies to increase performance in matching elements and materials from a BIM model to a knowledge database to enrich environmental indicators of commonly used elements' materials. The knowledge database contains all missing information for LCAs and has different levels of information for a range of several potential design options of elements and materials, including their dependencies. Accordingly, this paper investigates multiple NLP techniques and evaluates the performance of state-of-the-art deep learning models such as GermaNet, SpaCy, or BERT. Following this, the most performant NLP approach is used to provide an automatic workflow for matching Industry Foundation Classes (IFC) elements to the knowledge database, facilitating a seamless LCA in the early stages of design. For five different case studies, the performances of the proposed matching method are analyzed. Finally, one case study is selected to compare the embodied emissions results to those of the conventional process.