Challenge 2
Agentic AI for Automated CAD Generation and Autonomous Simulation _ SIAD Group
Team:
Aalen University of Applied Sciences – Human in Command Research Group (coordinator), with Starwit Technologies GmbH, Germany
Proposal title:
M3L Research Group, Institute of Mechanics SNCO (State Non-Commercial Organization), Armenia
Organisation
The Institute of Mechanics (Yerevan, Armenia) is one of Armenia’s leading research centres in solid and fluid mechanics, with roots dating back to 1955. Building on this foundation, the Mathematical Modelling and Machine Learning (M3L) Research Group was established in 2025 to connect advances in AI with engineering research and practice. The group combines expertise in computational mechanics, 3D computer vision, machine learning and agentic AI to develop methods for 3D reconstruction, automated CAD generation and engineering simulation.
Project info
The M3L team addresses SIAD’s routing bottleneck with an agent that writes compressor-skid piping as parametric CAD code. It builds on specialised CAD generators co-authored by its lead researcher (cadrille, ICLR 2026; CAD-Recode and CAD-Assistant, ICCV 2025), so candidate geometry is valid CAD from the start and the agent’s job is to correct it rather than invent it.
The system works in two regimes. In reconstruction, where a dimensioned arrangement drawing exists, the agent renders its model into the drawing’s projection, measures discrepancies in millimetres symbol by symbol, and rewrites the code until it matches. In design, where no such drawing exists, the same measurement tools score candidate routings from the P&ID, the CAD assembly and natural-language requirements against topology, clearance, interference, bend radii, supports, standards, and piping length and envelope.
Reconstruction across SIAD’s historical drawing–model pairs mines implicit conventions into a machine-checkable constraint set, which then verifies new designs and provides a reward signal for reinforcement-learning fine-tuning without labelled data. Because geometry is code, fittings, welds and bends are named features that drive mesh seeding and automated convergence studies. Final approval stays with the engineer.
Details
Routing compressor tubes inside a skid depends on scarce senior engineers who must interpret 2D drawings, adapt to existing assemblies and apply conventions that are rarely written down. SIAD’s historical data pairs drawings with 3D models but carries no labels saying which designs are good, which makes conventional supervised learning hard to apply.
The M3L Research Group at the Institute of Mechanics in Yerevan proposes an agentic system that generates piping as executable parametric CAD code, starting from CAD-generation models published by the team. The method uses SIAD’s drawings as supervision: by reconstructing historical cases and measuring every deviation in millimetres, the system learns the implicit rules behind them, including clearances, bend radii, spacing, support habits and routing idioms. That learned constraint set then verifies new routings when no drawing exists and drives reinforcement-learning fine-tuning. The same parametric model names each fitting, weld, flange and bend, so mesh seeding and convergence studies can be automated from the geometry itself. Engineers keep final approval.
The project will take the integrated system from TRL 2 to TRL 5, validated on unseen cases inside SIAD’s review workflow. Targets are an engineering-effort ratio of 0.5 or less on reconstruction cases, main piping length cut to 95% or less of the baseline where a compliant shorter route exists, and automated simulation results within 5% of engineer-accepted references in half the manual time. Shorter piping reduces steel, welding and shipping volume, and the reusable constraint-mining approach applies to other sectors where 2D drawings and 3D CAD coexist.
