Overview

This talk from the Autonomous Industrial Systems Laboratory (AISL) explores how to increase autonomy in process industries by combining situational awareness with intelligent decision-making.

We introduce a new framework for safe learning and optimisation that moves beyond traditional methods by using Gaussian-process surrogates to guide safe, guided exploration while optimising for key performance indicators like quality, throughput, and energy efficiency. An industrial case study illustrates the architecture, shows how embedding domain knowledge stabilises early learning, and outlines adversarial robust optimisation to maintain performance under disturbances.

We then discuss how AI can autonomously build models required for modern control, comparing fully data-driven and gray-box (physics+data) identification and highlighting automated model selection. Finally, we present an agentic workflow for abnormal situations: a Large Language Model (LLM) proposes actions, a digital twin agent validates them, and the LLM agent iteratively refines the plan in a closed loop.

Together, through interpretable state estimation, safe learning layered over existing control systems, and human–AI collaboration, we outline a practical path to autonomy that improves performance and resilience without compromising safety.

Speaker

Javal Vyas, Researcher, Autonomous Industrial Systems Lab, Imperial College London

Javal is a researcher at the Autonomous Industrial Systems Lab, Imperial College London, co‑supervised by Professor Mehmet Mercangöz. He holds an MSc in artificial intelligence engineering (chemical engineering focus) from Carnegie Mellon University, where his thesis centred on mathematical modelling, optimisation, and ML-driven scheduling systems. His work at Imperial continues this trajectory, integrating multi-agent AI systems and LLM control agents with digital twins to enable autonomous, fault-aware industrial operations.

The material presented has not been peer-reviewed. Any opinions are the presenters' own and do not necessarily represent those of IChemE or the Process Management and Control Special Interest Group. The information is given in good faith but without any liability on the part of IChemE.

Webinar slides