An Interpreter Layer for AEC: Cross-Modal Alignment, Schema Mapping, and Compliance

Architecture, Engineering, and Construction (AEC) projects require rigorous coordination across diverse stakeholders and heterogeneous data formats. While industry standards emphasise digital interoperability (e.g., OpenBIM/IFC), day-to-day coordination still relies on manual interpretive labour: practitioners translate unstructured site evidence, such as photos, chats, and floorplan patches, into model-linked, schema-compliant records. This translation is error-prone, leading to information loss across handoffs, ambiguous evidence-to-element linkages, and non-compliance with regulatory schemas. Consequently, this thesis addresses the following core research question: How can AI act as an interpreter middleware to reliably align unstructured site evidence with project data in AEC workflows? We decompose this challenge into two specific research questions: (RQ1) Multimodal Grounding under Ambiguity: How can multimodal site evidence be reliably grounded to specific digital project elements (e.g., IFC entities) amidst geometric repetition and noisy data? and (RQ2) Ontology-Driven Retrieval: How can domain-specific ontological constraints (e.g., IFC type hierarchies and spatial relationships) transform probabilistic neural outputs into ontology-compliant retrieval results, thereby mitigating hallucination risks? To answer these, we propose a hierarchical neuro-symbolic architecture, includes a fine-tuned vision-language model extracts typed spatial constraints from mixed site evidence; and a symbolic layer executes a priority-ordered graph traversal over an enriched IFC knowledge graph, built from topology-mined training data, while a Graph-RAG reranker identifies the final ranked candidates. The thesis contributes an empirically validated framework for neuro-symbolic interpreter layer design in IFC-centric AEC workflows; a reproducible evaluation protocol that separates extraction quality from symbolic retrieval quality; and a concrete analysis highlighting where utilising richer graph fingerprints offers the largest unrealised retrieval gains.