Semantic enrichment for BIM-based building energy performance simulations using semantic textual similarity and fine-tuning multilingual LLM

To achieve the global targets of the Paris Agreement of limiting global warming, it is necessary to reduce the operational energy of buildings, which are responsible for around 30% of the global greenhouse gas emissions. Building Energy Performance Simulation (BEPS) is an established method to estimate the building's energy demand in early design stages. Building Information Models (BIM) provides geometric and semantic information to create precise Building Energy Models (BEM) in early design stages. However, manual enrichment of missing semantic information is still a time-consuming and laborious process. Therefore, we propose a novel methodology to automatically enrich missing information to BIM using Semantic Textual Similarity (STS) and fine-tuned Large Language Models (LLM). For every IfcSpace, we match room-specific space types and constructions with missing thermal properties using the semantic most similar pairs of the BIM model and the according databases. We use three real-world case studies to fine-tune LLMs, and two case studies evaluate the whole methodology. Different fine-tuning strategies, such as using different loss functions, adding opposing word pairs or domain-specific abbreviations, significantly improve the accuracy of the matching. At the same time, however, findings show that semantic matching based on multilingual fine-tuned LLM performs worse than translated, monolingually fine-tuned LLM. Finally, BEPS results from automatically enriched BEM only slightly deviate from manually enriched BEM.