Automated conversion of unstructured mesh models to structured parametric CAD models for architectural applications

Parametric models are foundational to Building Information Modeling (BIM) and digital construction, enabling efficient design, analysis, and lifecycle management. However, most 3D building models reconstructed from real-world data sources — such as 3D scanning, photogrammetry, or format conversion — are represented as unstructured mesh models. Although these meshes accurately capture geometric shape, they typically lack the parametric constraints, design intent, and procedural history inherent to native BIM objects. This semantic deficiency severely limits their direct applicability in workflows that require model editing, parametric adjustment, or semantic interpretation. To address this challenge, this paper presents an automated framework for semantic lifting of unstructured architectural mesh models into structured, procedural parametric CAD models. The proposed pipeline first analyzes the geometric and topological characteristics of the input mesh and employs a deep learning-based module to predict a plausible sequence of CAD modeling operations. Subsequently, a 2D parametric profile abstraction and parameterization module processes the geometric regions involved in the predicted operations, transforming them into constrained and editable 2D parametric profiles composed of basic geometric primitives, such as lines, arcs, and circles. Finally, a parameterization-driven reconstruction module executes the recovered operation sequence and parametric planar profiles within a CAD environment to procedurally regenerate a structured solid model with full parametric editability. Extensive experiments on diverse architectural datasets demonstrate the robustness of the proposed approach, showing that it can generate parametric CAD models that closely match the original mesh geometry while supporting reliable parameter-driven editing. This research provides an effective pathway for the semantic lifting of mesh-based building models, enabling their reuse in BIM reconstruction, digital twins, and architectural analysis applications.