AIChE Journal · 2026
ChatP&ID: talking to engineering diagrams
LLM agents answer questions about P&IDs by querying DEXPI-based knowledge graphs, reaching 91% accuracy at about $0.004 per query and using 85% fewer tokens than reading the raw files.
Process Intelligence Research · TU DelftAI that reads, designs and checks chemical plants
We combine machine learning with chemical engineering knowledge, from engineering diagrams and process design to safety and operation, and develop it together with industry.
Citations and h-index from Google Scholar, updated 3 October 2026.
Industry partners
- Fluor
Recognition for our team
- NWO Veni grant2023
- IChemE Senior Moulton Medal2024
- AIChE W. David Smith, Jr. Graduate Publication Award2024
- TU Delft Open Science Award2025
- Marie Skłodowska-Curie Fellowship, Ulderico Di Caprio2026
- NWGD Best Presentation Award, Max Theisen2026
- Subject Editor, Chemical Engineering Research and Design
- Vice Chair, DEXPI Scientific Advisory Board
Featured work
Selected results
Recent work from the group, from new methods to tools that are ready to be used.
Nature Chemical Engineering · 2024
Generative AI in chemical engineering
A perspective on how generative AI can design processes, create engineering documents and support safety studies, and on the research needed to get there.
Ongoing projectwith Danone
Topology-aware AI for spray drying
Graph neural networks that embed the topology of a multistage spray dryer to model its dynamics, as a basis for soft sensors, forecasting and control. Presented with the NWGD Best Presentation Award 2026.
Research
Research themes
We develop AI methods that build on chemical engineering knowledge, from molecules to plants. Our work is organized in seven themes.
Generative AI for engineering diagrams
Digitization, autocompletion and autocorrection of P&IDs and PFDs, Smart P&IDs based on the DEXPI standard, and ChatP&ID.
AI for process safety
AI support for hazard and operability (HAZOP) studies, built on machine-readable engineering diagrams.
Hybrid and physics-constrained machine learning
Models that combine mechanistic knowledge with machine learning and respect physical and operational constraints.
Optimization with embedded machine learning
Deterministic global optimization of problems that contain trained neural networks, Gaussian processes or KANs.
Multi-agent and LLM systems for process design
Teams of collaborating LLM agents with specialized knowledge and tools for chemical engineering workflows.
Graph machine learning
Graph neural networks for molecular property prediction and topology-aware soft sensors for process plants.
Reinforcement learning for process synthesis
Agents that learn to design chemical process flowsheets by interacting with process simulators.
From research to practice
DigiCo: AI for Smart P&IDs
DigiCo (Digitization Companion) grew out of more than five years of our research on AI for engineering diagrams. It converts P&IDs into machine-readable Smart P&IDs stored in a knowledge graph, the basis for ChatP&ID, automated error correction and AI support for HAZOP studies.
DigiCo is prototype software. Several hundred users have signed up for early access, and a first beta release is planned for early 2027.
Open source
Open-source software
We release our methods as open-source software so that others can use and build on them.
- pyDEXPIPython implementation of the DEXPI standard for machine-readable P&IDs.
- SFILES 2.0Text-based representation of process flowsheets and P&IDs.
- ENFORCENeural networks that satisfy nonlinear equality and inequality constraints.
- reluMIPMixed-integer formulations of trained ReLU networks for optimization.
- MeLOnMachine learning models for deterministic global optimization.
People
Our team
We are PhD candidates, postdocs and visiting researchers in the Department of Chemical Engineering at TU Delft, led by Artur M. Schweidtmann.
