Structural Plan-to-Model Conversion with Deterministic Geometry and Guarded Agentic Vision-Language Refinement

August 19, 2026 | Source: arXiv AI

Tags: computer-vision, structural-engineering, VLM, finite-element, AEC, agentic-AI

A new framework converts structural engineering framing plan PDFs into editable finite-element model drafts without task-specific detector training. A deterministic geometry stage handles scale estimation and five entity classes; a guarded agentic VLM layer proposes corrections within explicit bounds. On a 50-plan held-out benchmark: 92.2% column recall, 88.6% beam recall, 100% for walls and braces.

Details

Converting structural framing plans to editable finite-element model drafts is manual, slow, and error-prone. Existing drawing-understanding systems require task-specific neural detector training; existing LLM agents in structural engineering work from text or model data rather than the drawings themselves. This paper from Mohammad Talebi-Kalaleh and Qipei Mei presents what the authors describe as the first framework applying an agentic vision-language layer to structural component detection and FEM model drafting from framing-plan PDFs, without any task-specific fine-tuning. The architecture has two stages. The deterministic stage extracts primitives, estimates scale by dimension-ratio consensus, recognizes five entity classes (columns, beams, walls, braces, openings) with a drafting grammar, and assembles an editable layout. The agentic stage uses a vision-language model to propose typed corrections, constrained by candidates from the deterministic stage, operation-specific admission tests, change-level review, and fail-closed transactions. On a 50-plan held-out benchmark (generated after all rule decisions were frozen): scale estimated within 0.1% of reference for every drawing. Recall/precision: columns 92.2%/99.7%, beams 88.6%/99.0%, walls 100%/100%, braces 100%/100%, openings 100%/96.4%. Important caveats: the held-out set shares a generator with the development set; the evaluation excludes independently drafted plans, raster evaluation, analytical connectivity, and solver validation.