Cracking control (FCC / steam cracker) with AI — severity, feedstock and margin on the same screen.
Optimal severity in an FCC or a steam cracker changes with every feedstock cargo — but the CoA arrives by email, the off-line lab result comes back two shifts later and the propylene price moves without telling the panel operator. iLEAN cross-checks DCS + CoA + lab + prices + the senior operator's notes, and hands the panel operator a reasoned recommendation before the curve drifts. iLEAN does not touch the control loop — the person decides.
The APR moves the curve. The reality that moves the curve arrives late and through informal channels.
The APR and the DCS are brilliant at holding the setpoint, but optimal severity changes with a reality they never see:
- The feedstock cargo CoA — density, sulfur, nitrogen, metals, hydrocarbon distribution. It arrives by email from the supplier at 8:00, when the riser is already processing that feedstock. The panel operator opens it when they can.
- The off-line lab — octane, contaminants, product quality. It comes back two shifts after the sample. By then, conditions have already changed.
- The spot price of propylene, ethylene, gasoline, naphtha. When the spread moves, optimal severity moves with it. That number lives on the trading desk; it never reaches the panel operator.
- Catalyst condition and the senior panel operator's know-how — an in-house lab spreadsheet, the shift lead's notes, 25 years of plant intuition. The veteran's island, the least protected of all.
The APR holds; the panel operator corrects; the committee reviews the next day. The classic system works 99% of the time. That other 1% is where the quarter's margin goes — and, in a cracking unit, one point of spread sustained is many millions a year.
iLEAN does not touch the control loop — it seals the cracks around the APR.
The problem in cracking is not that the DCS or the APR are bad: they are excellent. The problem is that the information that moves the optimal curve lives on islands the APR never sees, and it reaches the panel operator late. iLEAN acts as the putty that fills those gaps, without asking you to change your DCS, your APR or the critical control loop. The critical loop lives in ring 1 and nobody touches it.
Edge over selected visual inspection points. Connect captures CoA, lab, prices and the veteran's notes at second zero. The agent cross-checks against the DCS and leaves the panel operator a reasoned recommendation. The person decides and moves the setpoint — the agent does not.
The three iLEAN pieces applied to cracking:
- Edge — a vision-equipped terminal over the unit's visual inspection points (flare status, furnace flames, thermal leaks outdoors with an IR camera, reading old panels with no interface). Where the FCC does not allow a camera, Edge acts as a gateway to the OT network from a passive mailbox (ring 1), accepting no inbound connections.
- Connect — the star piece here. It captures the cargo CoA that arrives by email from the supplier, the off-line lab result, spot product prices from an external feed, the panel operator's notes and decisions by voice/earpiece, the text message from the operations manager about the incident that never made it into the DCS. All of it at second zero, with no forwarding.
- Agent — cross-checks the state of the DCS against CoA + lab + prices + notes + catalyst condition, and hands the panel operator a reasoned recommendation on severity: “with this feed density and this propylene-naphtha spread, the optimum is 4 °C above the current setpoint, and here is the expected curve”. The panel operator decides. The agent never moves the setpoint. The three rings guarantee that the critical OT network is sacred.
Classic FCC operation vs. cross-checked operation with iLEAN
| Aspect | APR + DCS alone + scattered emails | With iLEAN Connect + Agent around the APR |
|---|---|---|
| Feedstock CoA | Supplier email at 8:00, the panel operator opens it when they can | Captured at second zero, the agent cross-checks it with the DCS |
| Off-line lab | Two shifts behind | Connect picks it up instantly and links it to the actual cargo |
| Propylene-naphtha-ethylene spot price | Trading desk; reaches the panel operator late | External feed captured by Connect, cross-checked by the agent |
| Senior panel operator's know-how | In their head; it leaves at retirement | Connect captures notes and decisions; it stays as an asset |
| Severity recommendation | Shift meeting, intuition | Reasoned, with expected curve and margins; the person decides |
| Critical control loop | APR / DCS, ring 1, sacred | Untouched — ring 1 stays intact |
Impact estimate for your plant — to be validated with your numbers.
The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.
- FCC or steam cracker unit with a modern DCS + an APR up and running + an off-line lab running 1-2 shifts behind + variable feedstock (2-3 types alternating) + an optimization manager working between the plant and trading.
- Pilot centered on Connect (CoA + lab + prices + the panel operator's voice/earpiece notes) + a severity recommendation agent that never moves a setpoint. Complementary Edge over visual inspection of the flare/furnaces. First value expected in a few weeks.
- Indicative payback between 4 and 9 months, depending on how sensitive the unit is to variable feedstock, how often severity deviations are caught late, and the volatility of the propylene-naphtha or ethylene-naphtha spread.
- The hard lever is one additional point of margin sustained on the spread, plus less catalyst scrap / coke thanks to anticipating drift. In a cracking unit, one point of spread held through the year pays for the pilot many times over.
- Verified external data for the committee: over 2000-2021, the most digitalized sectors improved productivity by up to 40%; the least digitalized, barely at all [1]. Deep digitalization separates those who survive from those who fall.
And the plant manager's reasonable doubt
“What if the AI recommends a severity that breaks the unit?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes data from several systems (cross-checking the CoA with the DCS and prices), the best models brought error below 1.5% [2]. And, decisively: the agent never moves a setpoint, ever. It only recommends to the panel operator. The three rings guarantee it: the critical loop lives in ring 1, sacred, with no inbound Internet. This is exactly what IEC 62443 and serious OT culture have always done.
[1] Fundación BBVA / Ivie — productivity and digitalization 2000-2021.
[2] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.
What people ask about cracking control with AI
What does “severity” mean in an FCC or steam cracker, and why does optimizing it matter?
Severity is the degree of cracking applied to the feedstock — governed by riser outlet temperature (FCC) or COT (Coil Outlet Temperature, steam cracker), residence time and the steam/HC ratio. More severity means more light olefins (ethylene, propylene) but more coke and more tube wear; less severity means more fuel but less margin. The optimal curve shifts with every feedstock cargo, and every APR (advanced process regulator) chases it. Optimizing it is the difference between selling your naphtha at gasoline price or as an olefin.
What does AI add on top of the APR / DCS we already have, without touching the control loop?
iLEAN does not touch the critical control loop — the APR and the DCS still own the setpoint, sitting in ring 1 (the OT network, no inbound Internet). iLEAN works in rings 2 and 3: it cross-checks the reality of the DCS with data the APR never sees (the feedstock cargo CoA, off-line quality lab results, spot product prices, the turnaround schedule, catalyst condition), and hands the panel operator a reasoned recommendation before the number has to be dealt with manually. The person decides and moves the setpoint — the agent does not. This architecture is exactly what the IRIS category demands: capturing reality from the bottom up without giving up OT security.
Which “unintegrated” data carries the most weight in FCC severity?
It is almost never the DCS data: it is the data that arrives late through informal channels. (1) The cargo CoA the supplier emails at 8:00 when the riser is already processing that feedstock. (2) The off-line lab results for octane, density and contaminants (S, N, metals) that come back two shifts later. (3) The text message from the operations manager about an incident that never made it into the DCS. (4) The catalyst condition (activity, fines, metals) sitting in an in-house lab spreadsheet. iLEAN Connect captures all four at second zero and hands them to the agent that cross-checks them against the DCS.
Can a cracking plant trust an AI when its whole history is human operation?
That is exactly the right question — and the answer is that properly framed industrial AI learns from the senior operator, it does not replace them. The panel operator with 25 years on the FCC carries a mental map of how the severity curve responds to each feedstock; that map is in their head and it leaves with them. iLEAN Connect captures that know-how as data (notes, decisions, corrections) and the Agents use it as a reference. The augmented plant — the senior operator's judgment multiplied across all three shifts, instead of walking out the door at retirement.
How much does an AI pilot cost for an FCC or steam cracker unit?
The order of magnitude of a pilot on a cracking unit is close to that of a Connect+Agents pilot in a petrochemical plant with a modern DCS and an APR already running — the bottleneck is usually capturing the feedstock CoA + lab + prices + operating notes, not reading the DCS. A reasonable payback to put in front of the committee is several months — the hard lever is one additional point of margin on the propylene-naphtha or ethylene-naphtha spread, sustained over time. We ask for your plant's data and send you the estimated ROI in 48h.
Related: Batch reactor control with AI · Predictive maintenance for heat exchangers · CEMS emissions with AI.
Tell us your case and in 48h we'll send you the estimated ROI of this AI project for your cracking unit.
We work on your plant's real data, not ours. Diagnostic with no commitment.
Request estimated ROI in 48h See petrochemical