Feedstock switching has become a competitive necessity in petrochemicals, but capturing the margin opportunity without bleeding yield during transitions remains one of the hardest control problems in steam cracking. Transitions aren't a moment in time—they create a moving target for furnace severity, dilution steam, and coil balancing while coproduct impacts ripple across downstream units. Traditional APC models break down because feed composition changes shift cracking kinetics, coking tendency, and heat absorption simultaneously, and operators rationally revert to manual control when predictions diverge from reality. AI optimization learns coking behavior under specific feed blends and adjusts severity closer to the economic optimum during transitions, helping plants recover the yield that fixed conservative assumptions leave behind.
Every operations leader running a steam cracker knows the math changes when the feed changes. Switching from naphtha to ethane can push ethylene yield significantly higher, but it can also sharply reduce coproduct revenues for propylene and butadiene. The price difference between naphtha and ethane feedstocks has reached up to $400 per metric ton during periods of elevated oil prices.
That gap has made feedstock switching a competitive necessity, not a future ambition. Yet capturing that flexibility without bleeding yield during transitions remains one of the hardest process control constraints in petrochemicals.
Feedstock flexibility creates margin opportunity, but transition periods expose yield losses that traditional control systems struggle to prevent.
The sections below examine why feedstock transitions cost more than steady-state comparisons suggest and how AI optimization addresses the gaps.
Feedstock transitions cost margin because the transition itself is an operating regime, not a moment in time. The spreadsheet may show a lighter feed producing higher ethylene yield than a heavier one. What it rarely captures is the margin surrendered while the unit moves from one optimum to another.
In practice, many sites don't switch feedstocks with a clean step change. They blend in storage, ramp at the battery limits, or chase availability across pipeline batches and ship parcels. That creates a moving target for furnace severity, dilution steam, and coil balancing.
When the feed is changing faster than the plant can validate it, crews often choose stability first, and that yield giveaway can last longer than the planning model assumes.
Cracker residence times in the pyrolysis section are measured in tenths of a second, so small deviations in feed properties show up immediately in selectivity. Coil outlet temperature, steam-to-hydrocarbon ratios, and firing distribution all need coordinated adjustment, and the sequence matters. If firing is pushed before dilution steam and coil balancing catch up, tube metal temperatures can run up quickly. If dilution steam moves too early, the convection section and downstream compression see swings that create their own constraints.
The coproduct impact compounds the problem. Shifting to lighter feeds generally reduces propylene and C4 availability, which can pinch downstream units counting on those streams. A propylene shortfall can reduce load on a polymer unit, while a C4 shortfall can change extraction rates and inventory strategy for butadiene or raffinate.
Those effects show up as sitewide margin shifts, not just furnace economics. Planning teams working from static LP models may not capture these cascading effects until after the transition is underway.
Feedstock and derivative economics remain volatile across regions, and the value of flexibility increasingly depends on how well plants manage feedstock variability, not just on which feedstock is cheapest on paper.
Traditional advanced process control gets stressed during feed transitions for a straightforward reason: the control model was built around a narrower operating envelope than the transition actually creates.
When feed composition changes mid-run, operators often see it first in the divergence between what the APC predicts and what the analyzers report. APC models rely on linear dynamic relationships identified for a specific feedstock, and as composition shifts, the gap between model output and actual process behavior widens.
Steam cracking amplifies this because multiple relationships shift at once. Feed composition changes the furnace heat absorption profile, cracking kinetics, and coking tendency.
A coil outlet temperature move that produced a predictable ethylene improvement on one feed can produce a different response on another, especially when dilution steam, coil pressure drop, and firing distribution are also moving.
Measurement dynamics compound the control problem because many key quality signals are delayed. Online GCs update on cycles, lab confirmation arrives later, and inferential estimators trained on historical relationships that assume a narrower feed range can drift directionally wrong during a transition.
Plants compensate by de-tuning applications and operating with larger safety margins. That protects equipment and product specs, but it surrenders the yield optimization APC was deployed to deliver.
The root cause runs deeper than calibration. Heavier feeds introduce aromatics and higher molecular weight species that alter cracking kinetics and coke formation rates. A model tuned for light feed cracking will structurally mispredict behavior when processing heavier material.
That's an architectural mismatch between linear control logic and nonlinear chemistry, not a parameter adjustment problem. The fact that the industry is actively exploring AI approaches that learn nonlinear behavior tells you something about the limits of conventional APC architecture.
And that mismatch erodes operator trust at the worst possible moment: when model predictions diverge from what the board is showing, experienced operators rationally revert to manual control. No AI system replaces the pattern recognition that comes from decades at the board, but manual operation means the plant runs on instinct rather than on data-driven analytics.
Feedstock transitions are more often limited by coordination across teams than by any single unit. Operations sees real-time process data in the DCS, engineering works with simulation tools and historical data, and planning uses scheduling software with limited real-time feeds.
Each team makes decisions based on incomplete information, and transition decisions can still depend on manual handoffs because critical context sits scattered across tools and interfaces.
The coordination lag shows up first in constraints management. Operations may see early warnings: rising coil pressure drop, compressor approach to a limit, or a quality swing requiring severity pullback. Planning may still be operating on assumed yields and a transition duration that doesn't reflect what's happening in the control room.
Engineering may be evaluating longer-term constraints like fouling, run length risk, or equipment duty limits without a direct line to the transition choices being made hour to hour.
Severity adjustments during a transition also ripple beyond the furnace. A severity push that improves ethylene yield can create a heavier burden on hydrogenation, alter refrigeration duty, or change fractionation cut points. Those sequential dependencies, combined with disconnected information systems, create delays precisely when speed matters most.
The facilities that navigate transitions most effectively have something in common: a single model of plant behavior that operations, planning, and engineering can reference simultaneously.
When a planning team can see how proposed feed changes affect severity requirements and transition time in near real time, and engineering can validate implications before the transition begins rather than after, the coordination lag compresses.
That shared model also changes the tone of cross-functional conversations. Instead of debating whose data is "right," teams can align around the same predicted trade-offs, and that alignment often matters as much as the optimization itself.
The central operating decision during any feedstock transition is severity management under uncertainty. Higher coil outlet temperatures typically increase olefin yields but accelerate tube coking and shorten run length. Pulling severity back protects run length at the cost of immediate yield.
Transition periods make this trade-off sharper because the coking tendency of the new feed, and the actual cleanliness of the heater and convection section, aren't always known with confidence when the switch begins.
Traditional approaches often resolve the trade-off with fixed conservative assumptions. A site may carry a uniform safety margin on tube metal temperature or pressure drop because pushing into an unknown coking regime carries real metallurgical risk. That's also why the "best achievable" severity on steady-state operation can become "safe but suboptimal" during a transition.
AI optimization trained on plant operating data takes a different approach. Instead of assuming a static coking rate, it can learn how coking proxies move under specific feed blends and furnace conditions, then adjust severity to stay closer to the true economic optimum for longer stretches of the run.
During a feed change, live tube metal temperature modeling can maintain safe metallurgical limits while tightening the gap to the optimum. How quickly optimization adapts to changed feed properties determines how much yield a cracker recovers during a transition, and AI systems that learn continuously from operating data can adjust recommendations as the unit response changes rather than waiting for periodic model retuning.
An honest limitation: the model only sees what the instrumentation sees. If key sensors are unreliable, or if disturbances go unmeasured, the optimization needs tighter guardrails and more operator judgment. The implementations that hold up in operations treat AI as a decision tool inside an operating envelope, not as a replacement for metallurgical limits or process safety.
Implementations that work best often start in advisory mode, where the AI recommends setpoint adjustments and operators decide what to implement.
Advisory mode delivers value even if a site never moves beyond recommendations: it gives teams what-if analysis when constraints conflict (yield versus run length, ethylene versus coproducts) and a way to rehearse a feed transition before making it live.
Plants also use advisory recommendations to reduce cross-shift variability by giving each crew the same optimized starting point and a shared explanation of the trade-offs behind the move.
For operations leaders navigating feedstock volatility, the path forward requires optimization that adapts to changing feed composition without sacrificing yield or demanding manual intervention at every transition. Imubit's Closed Loop AI Optimization solution learns from actual plant operating data to build dynamic models that adjust setpoints in real time as feedstock properties change.
Plants can start in advisory mode to capture immediate value through decision support, scenario testing, and more consistent shift-to-shift execution, and then progress toward closed loop optimization as confidence builds. The result is faster, cleaner feedstock transitions with less yield giveaway and tighter coordination across functions.
Get a Plant Assessment to discover how AI optimization can reduce yield losses during feedstock transitions at your facility.
Feedstock switches reshape the entire product slate, not just ethylene output. Shifting from heavier feeds like naphtha to lighter feeds like ethane increases ethylene selectivity while reducing propylene and C4 coproduct availability, which can tighten constraints across downstream units. Teams evaluating flexibility typically make better decisions when they connect the transition plan to routing, storage, and downstream constraints through a broader feedstock optimization strategy, not only furnace economics.
Residence times on the order of tenths of a second mean feedstock composition changes create immediate yield consequences. AI optimization trained on plant data can detect when plant response no longer matches the prior feed assumption and recommend coordinated moves across severity, dilution steam, and constraints in near real time. The advantage is faster adaptation of targets and limits to match actual residence time distribution and process response.
Advisory mode keeps operators in control during transitions: the model recommends setpoint moves, and the board decides what to implement and when. That makes it practical to trial different severity ramps or steam ratios against metallurgical and downstream constraints without committing to automatic moves. Over time, crews see which recommendations hold up across shifts and feed blends, building trust and reducing variability. Sites that later choose closed loop AI carry the same boundaries and learnings forward.
