Challenge 2

Agentic AI for Automated CAD Generation and Autonomous Simulation _ SIAD Group

Team: 

DigitalTwin Technology GmbH, Germany

Proposal title: 

TRACE — Trustworthy Reasoning Agents for CAD & Engineering simulation

Organisation

DigitalTwin Technology GmbH is a Cologne-based SME, founded in 2018, that builds physics-based digital twins and simulation products for manufacturing, construction, automotive, semiconductor and ICT. It combines AI, finite-element simulation, IoT and 3D visualisation, and participates in 14 EU-funded research projects.

Project info

TRACE addresses the root cause of slow skid design: the knowledge needed for a validated piping design already exists, but it is scattered across 2D drawings, P&IDs, assemblies and standards, or held only in experts’ experience.

The system compiles drawings, CAD models, metadata and natural-language requirements into a formal Piping Constraint Graph in which every element records its source. Deterministic solvers then build all geometry from catalogue-based components and a routing optimiser with an explicit compactness objective, so constraints hold by construction rather than depending on an LLM writing correct geometry. A simulation agent automates meshing, convergence and critical-region detection, and its findings return to routing as new constraints, closing the loop between design and simulation. A rule-based critic checks outputs against EN/ASME standards before engineers review them at a built-in gate.

The workflow runs as a deterministic LangGraph state graph using self-hosted, open-weight models, so SIAD’s data stays on controlled EU infrastructure. SIAD receives parametric STEP models with editable parameter files plus a constraint-enriched 2D deliverable. Commitments include an engineering-effort ratio of 0.5 or less, smaller packages than the baseline, and at least 95% of outputs passing automated verification.

Details

Each new compressor skid requires a senior specialist to re-derive knowledge that already exists but is scattered across drawings, P&IDs, CAD assemblies, standards and personal experience. This step slows delivery, leads to oversized packages and avoidable piping cost, and makes simulation quality depend on who sets it up.

TRACE, developed by DigitalTwin Technology GmbH, turns design and simulation into a single traceable, closed-loop workflow. Vision and language models interpret the heterogeneous inputs and compile them into a Piping Constraint Graph; deterministic solvers then generate parametric, standards-compliant 3D piping optimised for compactness. A simulation agent automates meshing and convergence analysis, identifies critical regions and feeds its findings back into routing as new constraints. A standards checker and per-element provenance give reviewers evidence they can audit, while engineers keep final authority. All models are open-weight and self-hosted on EU infrastructure, so confidential data never leaves controlled environments.

The project will move the system from TRL 2–3 to TRL 5–6 through validation on unseen SIAD cases against expert baselines. Committed results are engineering effort at no more than half the manual baseline (with a stretch target of 30%), main piping length around 10% shorter, component and connection extraction F1 of at least 0.9, at least 95% of outputs passing automated verification, and simulation loops completing without intervention in at least 80% of cases. More compact skids reduce steel use, weight and transport footprint, and the architecture can be adapted to HVAC, hydraulic manifolds and shipbuilding piping.