Challenge 2

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

Team: 

DeepTech Chain LTD (team “PIPE-AGENT”), Cyprus

Proposal title: 

PIPE-AGENT — an agentic design-to-simulation assistant for compressor-skid piping

Organisation

DeepTech Chain LTD is a woman-led, research-driven SME based in Nicosia, Cyprus, founded in 2022. It designs and delivers deep-tech solutions that combine AI, robotics, IoT, cybersecurity and software engineering, and contributes as a technical partner to Horizon Europe and other EU collaborative innovation projects.

Project info

PIPE-AGENT automates the workflow SIAD’s engineers currently perform by hand when routing compressor tubes inside skids: reading 2D piping drawings and P&IDs, inferring implicit constraints, adapting components to existing assemblies, and setting up meshing and convergence studies.

Four cooperating agents share a machine-checkable constraint graph. A requirements agent, built on a fine-tuned open-weight multimodal LLM, turns natural-language requirements and drawings into typed constraints and flags uncertain extractions instead of guessing. A geometry agent generates editable parametric 3D piping on an open CAD kernel (CadQuery/Open CASCADE), solving routing as a constrained optimisation problem rather than by generative sampling. A simulation agent runs meshing, seed strategy and convergence analysis (Gmsh/CalculiX) and feeds critical stress regions back for refinement. A reviewer agent checks outputs against ASME B31.3 and EN 13480 rules and produces a traceability report linking every dimension to its source.

Engineers keep final approval. Where CAD-toolchain integration is not yet robust, the system delivers a constraint-annotated 2D output. Targets are an engineering-effort ratio of 0.30 or less and a package-size ratio of 0.90 or less against SIAD’s baseline, validated on unseen subsystems to reach TRL 5.

Details

Designing the piping of compressor skids still depends on senior engineers who interpret drawings, infer constraints that are never written down, and set up structural simulations by hand. These specialists are scarce across Europe’s process industries, and the manual workflow slows delivery and leaves layouts larger than they need to be.

PIPE-AGENT, developed by DeepTech Chain LTD for SIAD Group, is an agentic AI assistant that covers the path from design intent to simulation setup. Its central design choice is a strict division of labour: language models interpret requirements, P&IDs and 2D drawings, but geometry is only created if it satisfies a machine-checkable constraint graph covering connection points, diameters, clearances and piping-code rules. Routing is solved as a constrained optimisation problem, and a simulation agent runs meshing and convergence studies autonomously, returning stress hot-spots to the geometry stage. Every dimension can be traced to the requirement or drawing it came from, so engineers review the output instead of reconstructing it.

The project will take the integrated pipeline from TRL 3 to TRL 5 by validating it on unseen SIAD subsystems inside the company’s own review workflow. Expected results are engineering effort cut to 30% or less of the manual baseline, skid layouts at least 10% more compact, and constraint-extraction precision and recall of 90% or more. More compact skids also mean less steel and lower embodied carbon, and the open-source toolchain is designed to transfer to other skid and package builders without licence barriers.