TLDR: This research introduces an agentic AI framework that uses Large Language Models (LLMs) for both high-level fault recovery planning and low-level continuous process control in industrial automation. The framework employs a multi-agent system with validation and reprompting loops to refine LLM decisions. Case studies on Finite State Machine traversal and temperature control demonstrate that models like GPT-4o achieve high accuracy and robustness, outperforming open-source LLMs, and show the potential for LLMs to unify complex reasoning and real-time control, despite current limitations in physical modeling and instruction adherence.
Modern industrial operations, especially in chemical processes, are facing increasing complexity, coupled with workforce shortages and intricate fault scenarios. Traditional automation systems, while robust in well-defined settings, often fall short when dealing with the dynamic and unpredictable nature of today’s industrial challenges. This necessitates a new approach that can combine symbolic reasoning with adaptive control.
Researchers Javal Vyasa and Mehmet Mercangöz have introduced a unified agentic framework that leverages large language models (LLMs) to address these challenges. Their innovative system integrates LLMs for both discrete fault-recovery planning and continuous process control within a single, cohesive architecture.
How the Agentic Framework Works
The core of this framework is built upon Finite State Machines (FSMs), which serve as interpretable operating envelopes. The system comprises several key agents:
-
Monitoring Agent: This agent continuously monitors plant sensor streams and key performance indicators. It maintains the current FSM state and detects deviations, initiating the planning loop when necessary.
-
Action Agent: Upon detecting a deviation, this LLM-driven agent proposes a set of candidate control moves, which can be discrete transitions or continuous set-point adjustments, to guide the system back to its desired state.
-
Simulation (Digital Twin): Each proposed move is first executed offline in a high-fidelity digital twin of the plant. This allows for risk-free assessment of the action’s potential outcomes before real-world deployment.
-
Validation Agent: This agent scrutinizes the simulated outcome using rule-based checks and FSM semantics. If any limits are violated, it generates a structured rejection report.
-
Reprompting Agent: If a rejection occurs, this agent interprets the validator’s feedback and crafts a refined prompt for the Action Agent, highlighting violations and suggesting corrective heuristics. This iterative refinement process allows the LLM to improve its plan without human intervention.
This cycle of action, simulation, validation, and reprompting continues until a valid move is found or a predefined iteration limit is reached. If no satisfactory action emerges, a safety override hands control to a conservative fallback policy or a human operator, ensuring plant stability.
Case Study 1: Symbolic Recovery Planning via FSM Traversal
In the first case study, the framework was tested on its ability to generate valid recovery sequences within predefined FSMs. Fault handling was formalized as a path-planning problem, where the LLM agents had to navigate from a fault state to a normal operation state. FSMs were encoded as Python dictionaries for better interpretability by the LLMs.
The results were impressive, particularly with OpenAI’s models. GPT-4o and GPT-4o-mini achieved a 100% valid-path success rate across 180 randomly generated FSMs of varying sizes (4–25 states, 4–300 transitions) within five reprompts. They significantly outperformed open-source LLMs like LLaMA-3 in both accuracy and latency. While LLaMA struggled with instruction-following and multi-step symbolic planning, the OpenAI models completed planning cycles in under 4 seconds on average.
Case Study 2: Continuous Temperature Control
The second case study explored the framework’s capability in continuous process control. A dual-heater setup on a laboratory TCLab platform (and its digital twin) was used to maintain a target average temperature under persistent asymmetric disturbances. The LLM-based controller aimed to modulate dual-heater inputs to achieve this.
Compared to classical PID control, the LLM-based controller attained similar performance. GPT-3.5 operating on the digital twin showed performance closest to the PID baseline. GPT-4o demonstrated a more refined understanding of system dynamics, producing smoother control actions, though it required more reprompting. Locally hosted LLaMA models showed inferior performance but are attractive for offline deployment where low-latency inference and data privacy are prioritized.
The study also highlighted some reasoning limitations of LLMs, such as occasionally distributing equal power to heaters without accounting for their current temperatures, neglecting heat losses, or attempting to solve complex differential equations with overly simplistic steps. These observations emphasize the importance of grounding LLM predictions in validated simulation feedback.
Also Read:
- Designing Reliable Autonomous Systems: A Unified Approach to Control, Planning, and Learning
- AI and Digital Twins: Advancing Small Modular Reactor Technology
Future Outlook and Limitations
While the study demonstrates promising results, the authors acknowledge several limitations. The experiments were conducted on lab-scale systems with limited complexity, which may not fully capture the variability of real industrial environments. The reliance on commercial cloud-based LLM inference (e.g., OpenAI APIs) introduces latency, potential network failures, and recurring costs, which could be barriers for safety-critical industrial deployments where data sovereignty is crucial.
Future work will focus on deeper integration with existing industrial control hierarchies like SCADA and DCS systems. LLM agents may initially serve as supervisory advisors before being entrusted with full closed-loop authority. Incorporating domain-specific knowledge through fine-tuning or retrieval-augmented generation (RAG), coupling with physics-informed digital twins, and adopting standardized agent communication protocols are also key areas for development. Adherence to industrial standards for functional safety and cybersecurity will be critical for widespread adoption.
This research, detailed in the paper Autonomous Control Leveraging LLMs: An Agentic Framework for Next-Generation Industrial Automation, lays a significant foundation for LLM-based autonomous operators capable of reasoning, correcting, and adapting to novel situations in complex industrial environments, paving the way for resilient, language-driven automation in chemical engineering and beyond.
