Background

Detailed information on building elements, such as connection details or installation plans, is typically available in architecture and planning firms in an unstructured form, mostly as 2D drawings that are created manually for each project. 3D or BIM methods are rarely used, and standards for construction details are seldom applied. A systematic approach to capturing and reusing details is missing, which leads to high time requirements (3 to 8 hours per detail), inconsistent quality, and a loss of experiential knowledge. In light of increasing requirements for efficiency, carbon neutrality, and digitalization, the question arises how existing details can be digitally unlocked and integrated into BIM processes in order to preserve knowledge and promote sustainability.

Project Objectives

This exploratory study aims to technically and conceptually validate an AI-enabled workflow for digitally extracting and reusing construction 2D planning details as a knowledge resource for carbon-neutral, efficient planning processes in urban building construction. The study examines the viability of an approach that automatically detects existing 2D construction details, semantically classifies and parameterizes them, and provides the basis for generative and language-driven development of new details. This is intended to lay the knowledge foundation for an R&D project pipeline to develop a scalable, generative workflow for creating construction details based on individual (company-specific) project libraries.

Results

The result is a structured dataset with 20 to 30 typology-based reference details for timber construction and solid concrete construction in the residential and office sectors. A technology matrix was developed, and three AI architectures (CNN, ViT, GNN) were compared with regard to detection, segmentation, annotation, and generation of planning details, as well as criteria such as accuracy, training effort, and generalizability. In addition, a semantic model was created to represent building-component functions, construction principles, materials, and ecological aspects, including rules to ensure sustainability and circular economy principles in exemplary details. Usage scenarios and a UX concept for LLM-supported queries were prototypically demonstrated. This creates a solid basis for future collaborative R&D projects.