Coke accumulation in pyrolysis furnaces drives TMT rise, compresses severity windows, and creates fleet-level coordination constraints that traditional controllers struggle to manage. This article traces how coking reshapes individual furnace decisions, why pass imbalance shortens run lengths, and how multi-furnace dependencies exceed conventional control boundaries. AI optimization that models the full degradation cycle can coordinate fleet-level setpoints from advisory mode through closed loop control, extending run lengths and recovering margin across cracker complexes.
Every ethylene complex runs the same clock. From the moment a pyrolysis furnace returns to service after decoking, coke begins accumulating on radiant tube walls, tube metal temperature (TMT) starts climbing, and the window for maintaining target severity narrows. The decisions operators make during that window determine yield, energy consumption, and when the furnace comes offline again.
The stakes extend beyond a single unit. With European ethylene operating rates averaging just 70% to 75% in early 2024 and overcapacity pressuring margins globally, olefins producers need to extract more value from the assets they already have. Holding severity targets while managing coking rate across multiple furnaces is one of the clearest furnace energy efficiency levers available without new capital spending.
As coke builds through a run, operators lose room to maintain severity without pushing TMT toward its limit. Across a fleet, that becomes a shared throughput, energy, and coordination constraint.
The sections below trace how those constraints compound.
The fundamental operating principle of a pyrolysis furnace is straightforward: set feed rate, severity, and steam ratio so that all passes coke at the same rate and reach maximum TMT on the planned decoking date. In practice, that target keeps moving.
As coke deposits thicken on tube inner walls, heat transfer declines. The distributed control system (DCS) responds by demanding more fuel to maintain coil outlet temperature (COT). TMT rises as the insulating layer grows, and operators watch the gap between current temperature and the metallurgical limit narrow with each passing week. That narrowing accelerates: each additional increment of coke produces a proportionally larger TMT response, so operators lose headroom faster in the second half of a run than the first. Operators who extrapolate early-run TMT trends can underestimate how quickly that headroom disappears.
Those effects don't stop at the radiant section. Fouling in transfer line exchangers raises backpressure and increases compression energy demand downstream. Convection section fouling impairs feedstock preheating, and operators compensate by increasing firing further.
That additional fuel input accelerates thermal fatigue on tube materials, which can shorten tube life independent of the coking cycle itself. A furnace that started its run producing ethylene at target severity with moderate fuel per tonne ends its run consuming more fuel per tonne at lower selectivity.
TMT constraints become binding because coke accumulation reduces the achievable severity ceiling. The gap between planned and actual decoking date represents lost ethylene production that no downstream unit can recover.
A single pyrolysis furnace contains multiple parallel radiant coil passes. If one pass cokes faster than the others, operators may need to take the furnace offline for decoking before the remaining passes reach their full run potential. Every premature decoking event means lost production time and higher overall operating cost per tonne of ethylene.
The economics of decoking timing are asymmetric. Pulling a furnace offline too early leaves production capacity on the table: tubes that still had weeks of headroom contribute nothing during the decoke.
Running too long risks a forced shutdown from a TMT exceedance, or tube damage that extends the outage well beyond a standard decoking cycle. Operators typically err on the conservative side, which is the safer choice but means the fleet rarely runs at its theoretical maximum availability.
Equalizing cracking and coking rates across passes requires precise control of steam-to-hydrocarbon ratios and flow distribution to each pass. Even small imbalances create asymmetric coking patterns that shorten the run. The objective isn't simple volumetric balance but equalization of the coking rate itself: a quantity that can't be directly measured during normal operation. Tube skin thermocouples sample specific locations, and the actual hot spot on a given tube may sit between measurement points. That gap adds another layer of uncertainty to run-length decisions.
The difficulty increases with feedstock variability. A different naphtha composition can crack differently at the same COT. The yield distribution shifts and the coking rate changes, even without any change in operator-controlled variables. The same furnace running the same setpoints on Tuesday's naphtha may behave measurably differently from Monday's.
Firebox burner uniformity compounds the problem. Non-uniform burner performance causes localized overheating that accelerates carburization in specific tubes.
Firebox balancing is a tube life issue as much as a process control issue, and the interaction between burner condition and coking rate means that furnace maintenance decisions and process optimization decisions are never fully independent.
A typical ethylene complex operates multiple furnaces, with one usually in decoke, one in hot standby, and the remainder producing. Each producing furnace sits at a different point in its coking cycle, runs different feed compositions, and operates at a different severity. Combined cracked gas feeds shared separation and compression trains downstream.
This architecture means the plant is structurally never at steady state. Standard advanced process control and model predictive control architectures are typically designed around steadier assumptions.
A complex with furnaces in different stages of coking and variable-composition feedstocks shifts continuously. That reality is one reason plantwide process control has remained so difficult to achieve in olefins operations.
The measurement infrastructure reinforces the limitation. Severity isn't directly measured but inferred from product ratios or online analyzer outputs. Those analyzers are shared across several furnaces, so feedback arrives intermittently.
Between readings, controllers rely on interpolation and process model inference to estimate where each furnace actually sits.
Feed allocation introduces its own co-dependency. Altering feed selection changes fouling rate and run length, and when one furnace goes offline for decoking, the remaining furnaces must absorb lost throughput, which can push them harder and accelerate their own coking trajectories. Waves of new cracker capacity have reduced global ethylene utilization to roughly 80%, a trend that makes operational efficiency at existing assets even more critical.
Those variables don't move independently. Furnace severity affects downstream column loading, feed allocation drives coking rate, and co-product economics shift the optimal severity target.
No optimization model replaces decades of operating experience in a cracker complex, but the number of interacting variables across a furnace fleet exceeds what any operator or traditional controller can track simultaneously.
The degradation loop is a non-steady-state, multivariable problem. Coke accumulation raises TMT, increases energy demand, compresses severity, and destabilizes downstream operation at the same time. AI optimization is well suited to model that operating pattern because it learns from actual plant data rather than idealized assumptions about how the process should behave.
AI models trained on historical plant data can learn relationships between operating parameters and coking behavior for each furnace. A model that tracks coking state continuously can manage constraints as TMT headroom changes and coordinate fleet-level feed allocation decisions based on each unit's position in its cycle.
Advisory mode is usually where operators start testing the model against what they know. The model recommends setpoint changes, and operators decide whether to act. An operator who has managed a specific furnace through dozens of campaigns often has an intuitive sense for how that unit responds to feed changes or burner adjustments.
Advisory mode gives that operator a way to compare the model's predictions against lived experience. Where recommendations align, confidence builds quickly; where they diverge, the gap often surfaces assumptions worth examining on both sides.
Many plants start in advisory mode, move into supervised deployment as recommendations prove out, and progress to more automated control when confidence and operating readiness support it. That progression happens at each site's own pace.
That shared visibility matters during crew rotations too. When the model shows why one furnace can take more severity while another needs protection, operators at every experience level can see the coking-state reasoning behind each recommendation.
Operator training benefits from seeing optimization logic applied to real scenarios rather than classroom exercises. And because the model learns from actual operating outcomes, it preserves relationships between process states and the decisions that produced better results. That institutional record supports knowledge transfer as experienced operators move on.
For process industry leaders seeking to recover margin from existing cracker assets, Imubit's Closed Loop AI Optimization solution can learn from actual plant data across the full furnace cycle, write optimal setpoints to the control system in real time, and continuously adapt as coking state, feed composition, and market economics change. Plants can begin in advisory mode, move into supervised deployment as recommendations are validated, and progress to closed loop control as confidence builds.
Get a Plant Assessment to discover how AI optimization can extend furnace run lengths, reduce severity variance, and recover margin across your cracker complex.
As coke accumulates on radiant tube walls through a furnace campaign, TMT climbs and reduces the available headroom between current tube temperature and the metallurgical limit. Since raising COT to maintain severity also raises TMT, operators face a tightening window where target severity becomes unachievable without exceeding safe tube temperatures. Understanding where each furnace sits on that trajectory is the first step toward better throughput rate optimization.
Yes. Plant-wide optimization that tracks each furnace's coking state, current TMT trajectory, and remaining run-length headroom can help coordinate decoking timing across the fleet rather than treating each furnace independently. The objective is to keep decoking events staggered so plant-level throughput stays more consistent, which is one of the core benefits of self-optimizing plants in olefins operations.
Shared gas chromatographs limit severity control because severity is calculated from product ratios rather than measured directly at each furnace. When one analyzer serves several furnaces, feedback arrives only intermittently, and controllers must interpolate between readings while feed composition and coking state continue to change. Inferential models built on actual operating data can estimate severity from available process measurements and narrow the gap between analyzer timing and olefins plant optimization actions.
