Multi-agent systems for chemical engineering

We develop multi-agent systems based on large language models (LLMs) that split chemical engineering workflows into teams of collaborating agents, each with specialized knowledge and tools, and with the engineer in charge.

Challenge

Chemical engineering problems span multiple scales, from molecules to plant-wide operation and global supply chains, and they are solved by teams of experts who use specialized tools and information. LLM-based multi-agent systems are a young but fast-moving technology that mirrors this way of working. Early studies show promising results, but key scientific challenges remain: architectures tailored to engineering tasks, integration of heterogeneous data modalities (diagrams, simulation models, process data, text), foundation models with domain-specific modalities, and strategies that ensure transparency, safety, and low environmental impact.

Our approach

We see multi-agent systems in chemical engineering as interconnected, human-centric collaborators. Agents specialize in specific tasks while coordinating towards shared objectives. For example:

  • agents equipped with thermodynamic and kinetic models explore reaction pathways and materials at the molecular level;
  • at the process level, agents run simulation and optimization routines, balancing yield, energy efficiency, and safety;
  • at the plant and supply chain level, agents coordinate schedules, logistics, and inventories, while sustainability-focused agents evaluate environmental and economic performance.

Human engineers remain central. Multi-agent systems are designed to support and augment human expertise, not to replace it. This requires agents to communicate in ways that are transparent, interpretable, and aligned with their users, for example through natural language, engineering diagrams, modeling code, and experimental procedures.

The agents also need to be well integrated into the chemical engineering domain: interoperable with modeling and simulation tools, connected to validated databases, and able to interpret diverse data modalities. Our goal is not a generic end-to-end black-box AI system, but domain-specialized teams of agents. Building blocks from our group include machine-readable Smart P&IDs and ChatP&ID, LLM-generated reactor models (Text2Model), and AI support for HAZOP studies.

Key publications

  • Rupprecht, S., Gao, Q., Karia, T., & Schweidtmann, A. M. (2026). Multi-agent systems for chemical engineering: A review and perspective. Current Opinion in Chemical Engineering, 51, 101209. doi:10.1016/j.coche.2025.101209
  • Rupprecht, S., Hounat, Y., Kumar, M., Lastrucci, G., & Schweidtmann, A. M. (2025). Text2Model: Generating dynamic chemical reactor models using large language models (LLMs). Systems and Control Transactions (Proceedings of ESCAPE 35). doi:10.69997/sct.165009 · Details
  • Alimin, A. A., & Schweidtmann, A. M. (2026). GraphRAG for engineering diagrams: ChatP&ID enables LLM interaction with P&IDs. AIChE Journal, 72, e70540. doi:10.1002/aic.70540
← All projects