Inside an ethylene cracker, small efficiency losses build through shifting feed conditions, furnace limits and run-cycle progression. Those losses carry more economic weight in a downcycle, when margins are thin. The ethylene downcycle has put higher-cost capacity at risk of closure, and in that environment the efficiency gap between an average cracker and a well-optimized one can decide which assets stay competitive.
Feedstock economics set the floor, and operating discipline decides how much margin a plant captures above it. Feedstock positions are often fixed by geography and long-term contracts. The levers that matter most are the ones operations teams can move day to day.
How feed choice shapes cracker economics
The feed decision drives cracker economics more than almost any other variable. The gap between ethane and naphtha pricing has widened during recent market cycles. That gap explains why feed flexibility matters, and also why flexibility alone is not enough.
Naphtha-based operations produce coproducts such as propylene and butadiene that diversify revenue when those markets cooperate. Ethane-fed crackers run lighter and simpler, but carry more exposure to a single product and less flexibility when ethylene demand softens. A feed that looked best at last month's prices may not be the right call this week.
Crackers designed for several feeds, such as naphtha, propane and butane, can shift with pricing. Each feed change calls for different furnace severity, steam-to-hydrocarbon ratio and separation settings. A light feed such as ethane needs higher cracking severity to reach target conversion, while heavier feeds produce a wider coproduct slate that shifts load onto downstream fractionation. Getting those settings slightly wrong erodes the advantage that multi-feed capability was meant to provide.
A poor severity setting does not stop at ethylene yield in the furnace. It carries through to fractionation energy, product purity and compression load downstream.
Where efficiency erodes in cracker operations
Cracker efficiency erodes gradually across the run cycle, even when headline plant performance looks acceptable. The visible constraints usually show up earliest in the furnaces, in run length, and in the interaction between upstream and downstream units.
Furnace degradation and a shrinking operating window
Furnace heat transfer is one of the main operating constraints. As tube metallurgy ages and coke accumulates, heat transfer degrades through the run. Convection section fouling adds to the loss, sometimes unevenly across furnace passes. Plants often respond by backing away from the edge of the operating window.
That protects the equipment, but gives up yield and throughput that a healthier furnace could sustain, as covered under cracking furnace efficiency. When furnaces in a bank age at different rates, the constraint on one can force conservative operation across the whole bank, even when several furnaces still have headroom.
Coking, run length and decoke scheduling
Coking also creates uneven performance. When one coil or pass deteriorates faster than the rest, the weakest performer can end up setting the decoke schedule for the whole furnace. That shortens run length and reduces available throughput across the bank. The loss is rarely dramatic on a single day. It shows up as a series of small compromises that accumulate across the cycle.
Each run gets a little shorter, and decoking frequency climbs above what the original plan assumed. Over a year, those extra decoke cycles add up to real lost production time and work against olefins plant optimization.
Downstream separation and compression loads
Furnaces get most of the attention, but fractionation and compression loads stay tightly linked to upstream decisions. A change that looks favorable in the furnace can cost energy or separation balance downstream, which is why local optimization leaves value unrecovered.
Higher cracking severity, for example, produces more methane and hydrogen that must be compressed and separated before ethylene reaches the cold box. Plants rarely optimize those interactions as one system, and the cost of that fragmented approach is hard to see from any single unit's metrics. The operators running the furnaces and the operators running fractionation can each make good local decisions that add up to a weaker result for the plant.
The ceiling on static process control
Advanced process control (APC) has delivered real value in ethylene operations, particularly in stabilizing coil outlet temperatures and managing fuel flow. The same architecture also limits what APC captures when conditions keep changing.
Conventional APC holds a target and adjusts around it. It is less effective at rebalancing operating choices as the process moves through feed transitions, run-cycle progression or grade changes. The models behind it are usually linearized around a narrow operating region. They work well while conditions stay near that region and lose accuracy as the process drifts, so the gap between the controller's design basis and current conditions widens over the cycle.
What conservative operation costs
The usual result is a conservative operating posture. Plants trade yield for run length, and experienced operators compensate around the controller by adjusting setpoints by hand from pattern recognition. That works, but it does not carry across shifts or units, and the control layer captures less economic value the further the process moves from its original basis.
Operator judgment, built over years at the board, remains essential. Static linear models were built to hold key variables near their targets. Continuously finding the best economic operating point across a whole unit was outside their design. In practice, stability and run length take priority and some yield and energy opportunity is left behind.
How a learned model recovers cracker efficiency
A learned approach starts from the plant's own data. Dynamic process models trained on the cracker's operating history learn relationships as they actually exist on that unit, which can differ from what was expected at design conditions. A controller trained by reinforcement learning can then move several setpoints together as conditions change.
Coordinating across the full unit
Furnace optimization is often the highest-value starting point. Instead of treating control loops independently, the controller can coordinate moves across the furnace bank and account for how local decisions affect the rest of the unit. Many of the losses come from interactions between variables, so the wider view is what recovers them.
Adjusting severity on one furnace while accounting for its neighbors' coking state, the downstream fractionation load and the current feed composition gives a different answer than optimizing each in isolation. The same view applies to yield and energy, with setpoints across furnaces, fractionation and compression coordinated together.
Advisory mode as the starting point
Many plants start in advisory mode. The model recommends setpoint changes and operators decide whether to accept them. That lets teams compare trade-offs before committing, improve consistency across shifts and track how process behavior changes over time. Planning and economics teams can use the same model to run scenarios, test changing constraints and challenge LP assumptions built on older models.
Trust builds during this period. When experienced operators recognize their own decision patterns in the model's behavior, the system becomes practical to use. It supports their judgment, and the shared model gives operations and engineering a common reference for how the plant behaves today, instead of design-basis assumptions that may no longer apply. As recommendations prove out, the plant moves to closed loop, with operators setting the limits and able to take over from the board.
Common questions
How does feedstock variability affect the value of a learned model on a cracker?
Each change in feed composition shifts the best balance of severity, steam ratio, firing pattern and downstream separation load. Static control models tuned to a narrower basis struggle as conditions move through feed transitions and the run cycle. In plants that already have multi-feed flexibility, a learned model lets teams adjust continuously instead of relying on fixed targets that fit only part of the run, which is the basis of a self-optimizing petrochemical plant.
Can a learned model extend furnace run length between decokes?
It can help, by changing how the plant responds as furnace conditions evolve, especially when individual coils or passes deteriorate faster than the rest. Because the model reflects behavior across the furnace bank instead of isolated loops, operators get an earlier read on severity, firing and throughput trade-offs, and a clearer view of how decoke timing affects available throughput.
How does this differ from upgrading an existing APC system?
APC still delivers value, but its linear models have limits in a nonlinear, time-varying process like steam cracking, especially when feed conditions and run-cycle progression keep changing. A learned model is built from existing plant data, follows the process as it moves, and coordinates variables across furnaces and downstream units as one system. It sits on top of the existing APC and DCS and writes setpoints through them, inside limits set by plant engineers.
How the Imubit Platform applies
Imubit applies Map, Model and Control to ethylene crackers as part of its work in chemicals and polymers. Map traces lost yield, run length and energy to the recurring decisions behind them, such as feed assignment, severity, firing and decoke timing. Model uses Deep Learning Process Control® (DLPC®) to learn the cracker's nonlinear behavior across furnaces, compression and fractionation from its own history, guided by first principles. Control takes those decisions to closed loop, writing setpoints to the existing APC and DCS, inside limits set by plant engineers.
