Every shift in an ethane cracker faces the same tension: push coil outlet temperature higher for ethylene yield, or ease back to extend run length before the next decoke. With ethane crackers depending almost entirely on ethylene revenue and no coproduct cushion, optimizing severity across the full decoking cycle—not just today's yield per pass—is the difference between competitive operation and margin erosion. Traditional APC drifts as fouling progresses and coil conditions diverge, widening the gap between achievable and actual performance. AI optimization trained on plant data forecasts tube metal temperature trajectories, coordinates constraints across the complex, and maximizes cumulative ethylene output across the entire run.
Every shift in an ethane cracker control room involves the same fundamental tension: push coil outlet temperature higher to maximize ethylene yield, or ease back to extend run length and delay the next decoking cycle. That tension has always existed. What's changed is the margin environment.
The feedstock cost differential between naphtha and ethane has reached up to $400/MT since 2021, a gap that gives ethane-based producers their cost advantage. At the same time, the global ethylene downcycle has driven curtailments and increased closure risk for higher-cost capacity.
With substantial new petrochemical capacity still entering the market, concentrated in China and the Middle East, the pressure is not temporary. For ethane cracker operations, extracting every available yield percentage point through olefins plant optimization has shifted from incremental margin improvement to competitive survival.
Ethane crackers face a tightening margin environment where traditional control approaches leave measurable room between current and achievable performance.
The sections below break down these dynamics and what changes when AI handles the multivariable complexity involved.
Coil outlet temperature is the primary yield lever under direct plant control. Ethane cracking furnaces commonly operate with COT in the 1,400–1,600°F range, and given ethane's 80–84% ethylene yield (mass basis), even small COT variations translate directly to margin. But higher COT also accelerates coke deposition, driving tube metal temperatures toward the 2,000–2,100°F range as coke insulates tube walls.
Unlike naphtha crackers, which offset margins with propylene, aromatics, and C4 coproducts, ethane crackers depend almost entirely on ethylene. That single-product dependence means feedstock variability hits harder: if severity isn't optimized, there's no coproduct revenue to compensate. An ethane cracker running consistently below its achievable yield range leaves margin on the table every operating hour.
The optimal severity target is not a fixed setpoint. It shifts with ethylene pricing, ethane costs, decoking expenses, and maintenance windows. Early in a run, when coke accumulation is low and TMT headroom is wide, higher severity costs relatively little in run length.
Late in the run, the same severity push accelerates TMT at a steeper rate and compresses the remaining production window. The question every shift faces is whether today's yield gain justifies tomorrow's shortened run.
Decoking economics are often treated as a simple trade: one more day of run length versus one more notch of severity. In practice, the cost stack includes lost ethylene production, steam and fuel consumption for the decoke itself, and added workload for maintenance teams.
Equipment life takes a hit, too. Thermal cycling during shutdown, decoke, and restart contributes to tube and refractory stress over time. Plants that run at higher severity for short-term yield can pay that back later as tighter metallurgical limits, higher inspection scope, or earlier tube replacements.
Over multiple cycles, cumulative stress can narrow the gap between a furnace's rated TMT limit and the effective limit that tube inspections support. That shrinking envelope constrains every future run.
When one furnace comes down early, feed gets redistributed and constraints move quickly. Limits can shift to compressors, quench systems, refrigeration, and fractionation. Product quality risk can also migrate to units that were stable an hour earlier.
Preheat train and convection section fouling reduce heat recovery and raise fuel firing requirements. Higher firing means higher OPEX and higher CO₂ emissions. The harder cost to manage is what higher firing does to TMT trajectory and run length, especially when fouling is uneven across the train.
Because fouling progression is neither uniform nor perfectly measurable with standard instrumentation, it increases uncertainty when pushing severity closer to limits. In many plants, that uncertainty turns into a conservative operating buffer that persists until the next decoke resets the furnace.
Tying severity decisions to their downstream energy impact, covered in more detail under furnace energy efficiency, is one way to quantify what that buffer actually costs.
The severity-run length trade-off demands precise, adaptive control. Traditional advanced process control wasn't designed for that.
Model accuracy in operational ethylene cracking APC implementations tends to degrade as the run progresses. Key APC model parameters drift as feed quality assumptions, fouling degree, and coil maldistribution go unrecalibrated. With varying degrees of fouling across individual coils, estimating the precise temperature profile of each furnace coil becomes unreliable through model-based approaches alone.
That kind of model error creates a cascade: controllers get de-tuned, analyzer feedback gets discounted, and soft constraints replace true boundaries. Operating points drift further from optimal than necessary.
The precision gap is quantifiable. Industry experience suggests COT variation of approximately ±5°C under traditional control, widening to ±10°C during large disturbances, compared with approximately ±1°C in more advanced implementations.
That precision difference translates directly to economics: tighter control can support higher average severity and longer average run length without violating constraints.
Beyond precision, the architecture itself is a mismatch. Cracking processes exhibit strongly coupled multivariable dynamics: furnace outlet temperatures affect downstream separation performance, and product quality depends simultaneously on residence time, temperature profiles, and feed composition.
Single-loop PID architectures cannot coordinate these interdependencies. Even the furnace-to-fractionation interface typically operates as two separate control domains rather than the integrated system that plantwide process control requires.
The core shift is in the objective function itself. Traditional control optimizes for today's yield per pass. AI trained on actual plant data optimizes for cumulative ethylene output across the entire decoking cycle, accounting for the severity cost that shows up as shortened run length. That distinction matters most when unit behavior shifts mid-run.
AI can model nonlinear relationships between COT, coke deposition rate, TMT progression, and downstream separation performance simultaneously.
Those models update as conditions change. Operations move from reactive adjustments to continuous process control. Because the model tracks behavior at the individual coil level rather than relying on furnace averages, it can detect when specific coils are approaching limits faster than others and adjust firing accordingly.
Instead of responding after TMT approaches its limit, the model forecasts TMT trajectories under current conditions and adjusts severity to maximize total ethylene production across the full run.
Constraint coordination makes the value concrete. A furnace might have room on TMT, but the complex might be limited by compressor head, refrigeration horsepower, demethanizer overhead loading, or C2 splitter capacity.
Pushing severity can also tighten acetylene or methane constraints that show up downstream before they look urgent at the furnace. AI optimization does not eliminate those limits.
It makes the trade-offs explicit and consistent: across multiple furnaces at different fouling states, across shifts where different crews face the same constraint picture, and across competing objectives where the model can quantify outcomes before moves are made.
When planning, operations, and engineering teams share a single model of plant behavior, the decision-making process becomes grounded in shared data.
These systems integrate with an existing distributed control system (DCS). They receive live data and write optimal setpoints through established control infrastructure. Human AI collaboration typically starts in advisory mode, where operators validate recommendations before any automated control moves are made.
For operations leaders managing ethane cracker economics through a prolonged downcycle, the question is how quickly optimization technology can close the gap between current performance and the yield, efficiency, and run length these assets were designed to deliver.
Imubit's Closed Loop AI Optimization solution addresses this directly: it learns from plant data, writes optimal setpoints in real time through existing control infrastructure, and progresses from advisory recommendations to closed loop control as operator confidence builds. Plants can start in advisory mode, validate the AI's judgment against experienced operators, and move toward closed loop operation at a pace that matches their organization's readiness.
Get a Plant Assessment to discover how AI optimization can maximize ethylene yield and run length across your cracker complex.
Optimal severity changes because coke accumulation alters heat transfer and TMT response as the run progresses, even with consistent ethane quality. The same COT target produces a different constraint picture late in the run than it did on day one. Ethylene pricing, downstream separation loading, and proximity to planned maintenance also shift the economics of push versus protect. Feed variability compounds these dynamics further, as explored in petrochemical feedstock optimization.
AI optimization is typically layered on top of existing advanced process control and the distributed control system rather than replacing either. The AI layer reads live operating data and writes optimized setpoints through the same pathways operators already use, while APC continues handling regulatory control and basic constraint management. More detail on how the layers coexist is in industrial process control.
Preheat train fouling lowers furnace inlet temperature, so firing rates often rise to hold COT. Higher firing increases thermal stress and can accelerate TMT rise, shortening run length between decokes. Fouling also increases uncertainty in severity decisions because heat recovery loss is rarely uniform across the train. More on how fouling affects the broader heat transfer picture is available under heat exchanger optimization.
