Artificial Intelligence in Process Control

Brian Neve explains how AI can assist process engineers without replacing core expertise

ARTIFICIAL intelligence is advancing rapidly. But adoption in process industries will be slower – and for good reason.

Modern AI is probabilistic, predicting the most likely outcome based on training data. Process control must be deterministic. Safety shutdowns, interlocks or PID loops demand certainty: a probabilistic guess is unacceptable when managing high-pressure hydrocarbons or exothermic reactions.

The industry’s conservatism is reinforced by practical concerns: many industrial control systems are air-gapped, and security, functional safety and regulatory requirements constrain cloud-based AI. But to dismiss AI entirely would be a mistake. The revolution is happening now – in engineering, support and optimisation. Engineers and companies that fail to integrate AI today risk falling behind tomorrow.

AI in operations: soft sensors, tuning, and support

While the engineering and construction sector transforms rapidly, operational deployment faces connectivity and safety barriers. Many control systems are, and should remain, disconnected from the internet. Edge AI – local computers running tailored models trained on plant data – will become essential.

Soft sensors will integrate mass and heat balances across process units. Consider the Texas City refinery explosion in March 2005, where a distillation tower overfilled with hydrocarbons, killing 15 workers. A modern AI model monitoring real-time balances could have detected discrepancies between level indicators and actual inventory, preventing the disaster at minimal cost.

AI will also assist control loop tuning. A local AI model can predict instability, suggesting retuning or adjusting non-linear model predictive controllers (MPC) as feed quality or economic conditions change.

AI will act as a “sidecar” in safety incidents, providing guidance without replacing the safety system. For example, if a cooling water pump fails, AI can advise operators on sequencing actions and priorities, moving them from reactive to proactive management.

The immediate revolution: engineering and projects

The most immediate, high-impact application of AI is not on the plant floor but in the offices of engineering and construction companies and end users. Plant projects, expansions and turnarounds remain document-centric: lost spreadsheets, mismatched tag lists, hours of hunting for the right calculation. AI shifts this to a data-centric model, improving efficiency and reducing errors.

Consider a valve replacement during a turnaround. In the traditional workflow, discovering that both the original and spare valve were sized for an unusual backflow condition triggers weeks of manual verification, redesign and procurement. Startup may be delayed by months, costing hundreds of thousands of pounds.

In a more integrated scenario, AI trained on historical plant data, design models and procurement history can flag the mismatch weeks in advance, propose dual-mode configurations or bypass arrangements, generate updated documentation, and simulate the new operation. Engineers retain control but time-consuming coordination is dramatically reduced.

The human element: engineers in the AI era

The integration of AI introduces a significant skills gap. Instrumentation and process control are only lightly covered in undergraduate degrees, with mastery gained through postgraduate study or on-the-job training.

To thrive in this AI-enhanced environment, engineers, operators, planners and maintenance teams must develop a new skill: prompt engineering. This is the art of framing questions to AI to extract valid, safe and useful insights. Engineers should be practising this now: asking AI to explain reasoning, simulate outcomes or explore operating plans.

For example:

• Design engineer: “What is non-linear model-based predictive control? How does it differ from linear MPC? Create a simple simulation to illustrate the difference.”
• Operator: “Why was feedstock rerouted on crude tower 1 to tank CT013?”
• AI response: “A tanker bound for Rotterdam was rerouted due to weather, requiring an empty high-sulphur crude tank urgently.”

There are risks. A “lazy engineer” may take AI recommendations at face value. In one case, an AI suggested replacing a thermocouple but the true cause was a sticky fuel valve. The plant restarted with the same fault, leading to a furnace fire.

The key principle remains: AI is an assistant, not a replacement. Engineers must challenge AI’s probabilistic outputs and apply domain expertise. Corporate standards and guard rails for AI use are essential and must evolve continuously.

Preserving AI knowledge for the domain era

The conservative nature of the process industry will slow autonomous agents but it will not halt AI adoption in design, maintenance and support. Bespoke foundation models are unlikely from the likes of Google or OpenAI.

Engineers who learn to prompt, query and validate AI models will define the next era of process control

Instead, large control vendors such as Honeywell and Emerson, or specialised consultancies, will lead development – ideally guided by major end users such as ExxonMobil.

The “group of experts” model will likely prevail for critical decisions. Multiple AI agents, trained on different data subsets, will provide opinions that are weighted before a final diagnostic recommendation.

The bottleneck is domain knowledge. As senior control engineers retire, the deep understanding of process dynamics required to train and validate AI models risks being lost. Now is the time to transfer, store and encode this wisdom.

Engineers and companies that embrace AI – learning to prompt, query and validate models – will define the next era of process control. Those who do not will be left managing the spreadsheets of the past.

---