Copilot for constraint-driven generation of architectural design evaluations and suggestions
Supervisors: Prof. Achim Menges and Prof. Dr. Thomas Wortmann
Advisors: Anni Dai and Dr.-Ing. Tobias Schwinn
Building design, engineering, and construction are highly regulated, requiring time-consuming checks and costly expert consultations, slowing design iterations and worsening the housing crisis. Architects often navigate conflicting advice from specialists, leading to revisions and delays. Occasionally, issues emerge after construction, causing costly setbacks and highlighting the need for earlier, integrated compliance checks. This research explores an AI-powered methodology for compliance checking of building models under German requirements, combined with design suggestions based on evaluation results. It supports architects in the early design stages by offering evaluations, suggestions, and adjustments to ensure code-compliant buildings. The study proposes a methodology combining a multi-agent system with large language models (LLMs) and Retrieval-Augmented Generation (RAG) to interpret regulatory texts and assess graph-based BIM data. Domain-specific Evaluation agents (e.g., fire safety, accessibility) analyse the design, while a Supervisor agent resolves possible conflicts between domains and synthesizes the results. The Suggestion agent proposes necessary design changes, and a Validation agent checks for compliance. The architect remains in control, using the system as a copilot. A Revit add-in for the AI copilot demonstrates real-world integration. User studies assessed system performance, while a case study explored its impact on architectural workflows. The system detected compliance issues and suggested design improvements. The interaction between the architect and copilot reveals a complementary relationship, where the copilot supports decision-making by providing evaluations and suggestions, while the architect remains in charge of final design decisions. This dynamic enables a blended co-design workflow, improving efficiency and better informing design decisions.
