Thermal cracking breaks heavy hydrocarbons with heat rather than catalyst. In a refinery it happens in visbreakers and delayed cokers, where a fired heater brings the residue feed up to cracking temperature. The same ideas carry over to the steam cracking furnaces of an olefins plant. All of these heaters are limited by coke that deposits on the inside of the furnace tubes.
Coke insulates the tube wall, so tube metal temperature climbs for the same heat duty. It also narrows the flow path, so pass pressure drop rises. When the hottest tube or the highest pressure drop reaches its limit, the furnace comes off line for a decoke. A premature decoke costs the lost production, the maintenance work and the thermal cycling of the tubes. Coke buildup remains one of the main barriers to consistent profitability in these energy-intensive units.
Conventional optimization tools tend to target yield or furnace efficiency, and rarely both together. Industrial AI models learned from historian data can manage both at once and act before fouling forces the decision. Six strategies follow, starting with the balance between furnace passes.
1. Balance furnace passes to prevent early fouling
Uneven flow or coil outlet temperature across furnace passes creates local hot spots where coke forms fastest. The hotter pass runs at a higher tube metal temperature, fouls faster, and reaches its pressure-drop limit while the other passes still have margin. Days between decokes is set by that worst pass: when it reaches its limit, the whole furnace has to come down.
A multivariable model that reads temperature, pressure drop and firing rate from the distributed control system (DCS) can learn the nonlinear coupling between passes. It then trims individual pass flows or firing to equalize heat duty. Single-loop PID controllers act on one pass at a time. A model that considers all passes together reduces the temperature spread between them.
Before relying on pass balancing:
- Verify flow orifice sizing and correct any mismatch between passes.
- Install or calibrate pressure-drop transmitters so the model sees the true pressure balance.
- Connect the model to the DCS with guardrails that respect existing safety and emissions limits.
With balanced passes, fouling progresses uniformly and predictably. That adds production days between decokes and protects tube integrity.
2. Detect coke formation early
Fouling rarely announces itself with a single alarm. It shows up as small shifts in tube metal temperature, pass pressure drop or steam-to-hydrocarbon ratio. Inferential models trained on these high-frequency signals learn the combinations that precede coke laydown. They can raise an alert before conventional thresholds would trigger, which turns an unplanned decoke into a planned one.
Training starts with several months of historian data, including pressure-drop trends, product composition and lab results, paired with timestamped decoke records. The decoke records give the model the outcome it needs to learn from.
Variable-importance rankings show which signals drive each alert. What-if simulation lets engineers test how a modest change in steam-to-hydrocarbon ratio changes the fouling rate before they make it. Combining the data-driven model with rule-based logic gives both high recall and reasoning that operators can follow. Maintenance becomes planned, and run length extends.
3. Take control adjustments to closed loop
Holding coil outlet temperature, steam ratio, pressure and feed flow on target at the same time is a constant balancing act. The variables are coupled, so a move on one pass ripples through the furnace and can shift product selectivity. In steam cracking furnaces the reaction network is highly nonlinear. A controller trained with reinforcement learning, against a model learned from the unit's historian data, finds control moves that raise margin while honoring safety and emissions constraints.
Deployment starts in advisory mode. The model runs alongside the DCS, and operators compare its recommendations with their own moves. Once its recommendations consistently do better than manual adjustment, operators grant it closed-loop authority. The controller then writes setpoints to the existing advanced process control (APC) or DCS minute by minute, within guardrails set by engineering standards, with operators able to take over at any time.
The controller shows variable-importance rankings and what-if views, so operators can see why each move is made. That transparency builds trust and develops skills. It also settles the usual tension between maximizing yield and saving energy without round-the-clock manual tuning.
4. Stabilize feed and cut-point transitions
Switching feed slates or tightening cut points upsets a cracking unit. Temperatures drift, pressures swing and coke builds faster than operators can respond. A model trained on past transitions adjusts coil outlet temperature, steam ratio and feed rate, and with them residence time, as the transition proceeds, to keep the furnace inside its operating window.
The model tracks the coupled variables continuously and responds to disturbances that are hard to hold with manual moves alone. It works through the existing advanced process control layer. The aim is less temperature variation through the transition and fewer of the pressure swings that force emergency rate cuts.
Stability has direct value. Smoother cut-point moves keep product on spec, protect run length and let the unit return to full rate sooner instead of waiting to settle. As feed changes become more frequent, the models adapt with them.
5. Make counterintuitive moves that slow fouling
Pushing furnace duty is the instinctive response to a daily yield target, but the extra heat accelerates coke growth and shortens the run. Doing a little less at the right moment can be worth more over the full run.
Using historian, lab and economic data, a model can trace how profit and fouling rate change with severity. It finds the point where more duty costs more in run length than it earns in yield. Past that point, a modest trim in coil outlet temperature or a temporary cut in feed rate can delay the next decoke.
Typical moves include:
- Easing duty during short-lived spikes in feed sulfur.
- Holding steam-to-hydrocarbon ratio steady instead of chasing lab noise.
- Reducing severity early in the run to preserve tube skin temperature margin for later.
Because each recommendation carries its economic effect, operators can see that it is a measured trade between throughput today and asset health over the run.
6. Capture and retain operator knowledge
Experienced board operators keep furnaces stable on judgment built over years. Retirements and turnover put that judgment at risk. A model trained on the unit's operating history captures how the furnace responded to the moves those operators made. Daily logs, event notes and shift comments add the context the historian lacks, such as why a move was made, and help turn it into structured knowledge that stays with the plant.
The model explains its recommendations through variable-importance charts and what-if views. Newer operators can see how a suggested move fits safety limits and economic targets. Recording each intervention and its outcome builds a working playbook that shortens onboarding and keeps practice consistent across shifts. The model keeps learning from new operating data, so good practice carries over as the crew changes.
Where to start
The six strategies can be adopted in stages, starting where lost run length costs the most and within the constraints of each unit. As in strategy 3, alerts and advisory recommendations come before closed-loop control.
For a thermal cracking heater in refining, Imubit's work follows Map, Model and Control. Map traces lost run length and rate to the recurring decisions behind them, such as pass balancing, firing and severity targets. Model uses Deep Learning Process Control® (DLPC®) to learn a nonlinear dynamic model of the heater from its own history, guided by first principles. Control takes those decisions to closed loop, writing setpoints to the existing APC or DCS minute by minute, inside limits set by plant engineers.
