Challenge 2
Agentic AI for Automated CAD Generation and Autonomous Simulation _ SIAD Group
Team:
ITECAM — Centro Tecnológico Industrial de Castilla-La Mancha, Spain
Proposal title:
TRAZA — Constraint-grounded agentic AI for piping design and simulation setup
Organisation
ITECAM is the industrial technology centre of Castilla-La Mancha, based in Tomelloso, Spain. Serving industry, particularly the metal sector, it combines an applied AI department with established design and structural-calculation teams and capabilities in digital manufacturing, simulation, materials, testing and the energy transition.
Project info
TRAZA treats skid piping as a judgement problem, not a drafting one, and makes the engineer’s implicit knowledge explicit. Language models read and formalise; a deterministic solver reasons in space. Every output is therefore traceable to the constraint that justifies it and the document it came from.
Four agents form the pipeline. An extraction agent builds a connection graph from 2D drawings and P&IDs, parsing CAD files directly and using an open-weight vision-language model for scanned sheets and a small detector trained on SIAD’s own symbol legend. A constraints agent turns specifications into a versioned rule catalogue drafted by an LLM and validated by an engineer, then applies the rules deterministically. A geometry agent places components with constraint programming and routes pipes with path-planning algorithms and collision checking, guided by preferences learned from SIAD’s approved models, producing editable STEP solids. A simulation agent uses a graph neural network surrogate to screen candidate routes for nozzle loads, flexibility and natural frequencies during optimisation.
Targets are an engineering-effort ratio of 0.40 or less, a size ratio of 0.95 or less, and rework at 30% or less of the manual baseline, measured against a documented 550-hour skid project.
Details
Routing compressor piping inside a skid is limited less by drawing speed than by engineering judgement. A senior engineer must reconcile P&ID connection logic, CAD envelopes, specifications, maintenance access and structural limits, much of which is never written down as rules. This scarce expertise sets the pace of the vendor workflow and keeps advanced piping design out of reach for many SMEs.
TRAZA, developed by ITECAM, is an agentic AI system that converts this implicit knowledge into explicit, machine-checkable constraints before any routing takes place. Language models are used only where they perform well, reading heterogeneous documents and drafting rules for engineers to validate, while deterministic solvers generate the geometry. Routing is guided by preferences learned from SIAD’s approved designs so that results resemble accepted practice, and a fast structural surrogate checks nozzle loads and natural frequencies during optimisation rather than after it. Every element of the output points back to its source drawing and rule, giving reviewers an auditable report instead of an opaque model.
ITECAM builds on tools it has already deployed: a hybrid extractor for 2D industrial diagrams and an optimiser that converts DXF plans into 3D models evaluated by a trained surrogate. The project targets TRL 5 for the structured 2D deliverable on unseen industrial cases and TRL 4 for the full 3D and simulation branch. Expected results are a 60% reduction in piping-design effort, rework of 30% or less of the manual baseline, and at least 5% smaller skid envelopes and piping length, with less steel, fewer welds and fewer redesign loops.
