How Industrial Processes Are Controlled Today — And Why It's Not Enough

The foundation of modern optimization in complex process industries — refineries, petrochemicals, or cement plants — is process modeling. The goal is to describe the interaction of hundreds of variables (temperature, pressure, flow rates, raw material composition) and find their optimal combination for maximum profit while adhering to safety and environmental limits.

Traditional tools for Advanced Process Control (APC) solve this problem by simplification. Engineers perform so-called step tests — they deliberately change one variable and observe how others react. From the collected data (typically over 1–2 weeks of operation), they construct a gain matrix that describes the relationships between variables. This model then serves as the basis for automatic control.

The problem is that APC models are inherently linear. They assume that the relationship between temperature and yield is a straight line — which is almost never true in real chemistry. “It's like trying to predict the weather based only on what you saw last week,” describes Greg White, a technical consulting specialist at Imubit. Real processes are nonlinear, dynamic, and influenced by factors that step tests do not capture.

AIO: When Neural Networks Read Data Instead of Textbooks

A new generation of tools for AI-based Closed-Loop Optimization (AIO) takes the opposite approach. Instead of starting with theoretical assumptions about how a process should work, it begins with the data itself — months to years of operational history with minute-level resolution. Deep neural networks then uncover the true relationships in the data, including those that an engineer, relying on textbooks, would never think to look for.

Only in the second step do domain experts come in, interpreting, verifying, and adding boundaries derived from fundamental physical laws. This order — data first, then physics — is crucial. It allows the model to uncover unexpected interactions without theoretical assumptions "checking them off" beforehand.

“If we allow neural networks to operate outside the framework of fundamental physical laws in the initial phase, unnecessary bias is removed, and models can uncover real existing relationships that traditional modeling overlooks,” explains White.

Crucially, AIO models do not operate as black boxes. The judgment of technical experts remains central — from setting expectations in gain matrices to controlling outputs. The technology acts as an amplifier of human expertise, not a replacement for it.

Hydrocracking: When AI Reveals What Engineers Overlooked

A real-world example demonstrates the power — and limits — of the AIO approach. A hydrocracking unit is one of the most complex pieces of equipment in a refinery. Under high pressure and temperature, it breaks down heavy hydrocarbons into lighter, more valuable products such as diesel and jet fuel.

For one particular unit, the AI-based model was expected to detect so-called overcracking — a phenomenon where, as temperature increases, diesel yield initially rises, but beyond a certain point, diesel molecules themselves begin to crack into less valuable gases (butane, propane). Traditional APC avoids this tipping point because it violates linear assumptions. An AIO model should identify it and navigate it safely.

However, reality was different. The model's output showed all gains as positive — with no visible tipping point. After a deeper analysis, it turned out that the problem was not with the model, but with the quality of historical data. The operation used data compression, which omitted intermediate samples and thus obscured dynamic responses. At the same time, there was no correction for distillation fraction boundaries in product flow rates.

Despite these limitations, the model proved its value. When engineers segmented historical data by pricing regimes (when diesel is more expensive and when jet fuel is), the model revealed that the existing operation unintentionally favored jet fuel at the expense of diesel — which, in a high-priced diesel regime, meant a loss of opportunities amounting to tens of thousands of dollars per day.

Why Operator Trust Determines Success

Even the best AI model is useless if operators turn it off and revert to manual control. This is a key lesson, according to Imubit, which has already completed 100+ closed-loop deployments. Its platform is currently used by refineries of companies such as Chevron, CITGO, Marathon, Preem, Hunt Refining, and cement plants like Ash Grove and Eagle Materials.

“About 50–60% of our senior control center operators will retire in the coming years,” says Bryan Cook, Senior Automation Engineer at Ash Grove Cement. “It's huge for us that we can train new people in an open loop and let them experiment and make mistakes without real consequences.”

A key tool for building trust are “what-if” simulations. Operators can play with the model — shifting constraints, switching price sets, simulating changes in manipulated variables — and immediately see the impact on the unit's objective function. When a simulation shows that reducing the reactor's WABT temperature will yield a $10,000 per day higher margin, the operator trusts it much more than if they just receive a table of numbers.

Czech and European Context

These topics are exceptionally relevant for Czech industry. The Czech Republic has one of the highest concentrations of processing industry in the EU — refineries in Litvínov and Kralupy, chemical plants in Lovosice and Neratovice, cement plants across the country. All these operations face the same challenges as their American counterparts: experienced operators retiring, pressure on margins, and the need to reduce emissions.

Furthermore, the European Union, through the AI Act and the Industrial Emissions Directive (IED), is creating regulatory pressure for transparency and process optimization. Solutions like AIO, which combine deep learning with transparent output verification, can help Czech companies meet stricter requirements without losing competitiveness.

Czech companies are actively engaging in industrial digitalization — the Czech AI Factory in Ostrava, launched in 2026, is one of the nodes of the European AI supercomputer network and can also serve as a technological base for industrial AI applications. The first Czech companies are already experimenting with machine learning-based predictive maintenance, but complex process optimization using deep neural networks is still in its early stages here.

What the Numbers Say

Imubit reports concrete results from its deployments: a 15–30% reduction in natural gas consumption, a $0.25 per barrel improvement in margins, and a 1–3% increase in yield. In an industry with single-digit percentage margins, this means the difference between profit and loss.

For context: a medium-sized refinery processing 100,000 barrels per day can gain over $9 million annually with a $0.25/barrel margin improvement. And that's from just one optimization intervention — without investing in new hardware.

This is not a futuristic vision. Imubit has 7 years of model operation and deployments in real-world operations across North America and Europe. Their customers report not only economic benefits but also a cultural shift — operators who have adopted AIO tools, according to Big West Oil, “started thinking outside their comfort zone” and competing among themselves for better economic results.

Data as a New Asset — With Caveats

However, the transition to AIO is not free. It requires high-quality data without compression and with sufficient resolution — which, for older operations, may mean an investment in modernizing historical data systems. As the hydrocracking unit case showed, data compression used to save storage in the last decade now actively harms model accuracy.

The good news is that with today's storage prices, turning off compression and archiving raw data is an economically trivial step that will pay off many times over in better predictions. The recommended procedure is: turn off compression, archive data in its original resolution. Future AI models will thank you.