A vacuum column can stay inside its formal constraints and still give away margin shift after shift. A small rise in overhead pressure raises the temperature the column needs for the same lift, and recoverable gas oil stays in the bottoms. Because the loss sits in how the unit is run, it can be recovered from the existing column.
Separation is one of the largest energy users in refining and chemical processing, and distillation accounts for most of it. Industry as a whole uses nearly 40% of global energy, according to the International Energy Agency. In a vacuum distillation unit (VDU), operators balance vacuum level, temperature, steam rates and internal losses every shift as the feed changes, and the difference between the best operating point and an acceptable one adds up over every hour of operation.
The problem is rarely a single variable drifting out of bounds. More often the operating window moves while the targets stay where they were set. Many vacuum columns run well inside their real constraint boundaries because the interacting variables move faster than conventional control can coordinate them.
Pressure is the main lever
Vacuum distillation rests on one relationship: lower pressure lowers the temperature at which hydrocarbons boil. That matters in heavy service because above the thermal limit, valuable material cracks into lower-value byproducts and coke. A modest reduction in overhead pressure can improve recovery by lifting more gas oil out of the bottoms, and at typical throughput that yield shift changes the unit's annual economics. A small pressure penalty at the top of the column usually shows up later as lost recovery at the bottom.
The best overhead pressure is not fixed. It changes with feed rate, molecular weight, steam injection rate and current product economics. Operators work inside that moving target every shift, but in a unit where the vacuum system, column internals and feed heater all interact, the best target at 6 a.m. may not be the best target at noon. The gap between the best pressure setpoint and the one the controller is holding often goes unnoticed until a yield calculation shows what was left behind.
The narrow thermal envelope
Higher temperature lifts more heavy material into product. Push too far, though, and vapor velocity rises until heavy material is carried up into the gas oil and contaminates it. At that point the margin reverses. The envelope between too cold and too hot is narrow in vacuum service, and it moves with every change in feed quality or vacuum system performance.
Running a VDU efficiently depends on knowing where that envelope sits under current conditions, which the last calibration may no longer describe. Tracking it belongs with the other operational efficiency measures on the unit.
Equipment that shapes separation performance
Column internals
Column internals decide how much of the vacuum system's benefit reaches the separation zone. Every increment of pressure drop across the internals works against the vacuum system and raises the temperature at the base of the column. That is why structured packing is widely used in vacuum service: it gives lower pressure drop than trays for the same separation duty.
The advantage comes with trade-offs. Fouling tendency depends on the service and the internal design. In heavily fouling sections, open designs such as grid packing are often preferred, because structured packing would lose capacity between cleanings.
Ejector and condenser degradation
The vacuum system adds its own constraint. Steam ejectors are designed for efficiency at a single inlet pressure. In multistage ejector systems, condensers between stages remove motive steam and condensable vapors, and each downstream stage handles the remaining non-condensables at a higher pressure.
Condenser fouling is a common failure path. The decline is usually gradual enough that operators do not notice the operating envelope narrowing until a process upset exposes it.
When condenser or ejector performance deteriorates, the effect reaches the separation itself. Operators lose room to trade pressure, steam and temperature against each other. In deep-cut vacuum service, where the column is designed to recover as much gas oil from residue as possible (see Hydrocarbon Processing's design series on deep-cut vacuum units), this narrowing cuts directly into the most valuable product stream.
Pressure measurement accuracy
Pressure measurement can distort decisions too. In vacuum service, transmitter diaphragms can be damaged during startup, and the damage leaves a persistent measurement error in normal operation. Even a small error changes how operators judge their room, and it carries into temperature targets, steam rates and constraint limits at the same time. A bad reading can make a healthy move look unsafe, or make a constrained unit appear to have room it no longer has.
Any controller or model that acts on those signals inherits the error. Validating vacuum measurements is part of managing the risk of AI in process operations.
Why fixed control strategies leave margin behind
Conventional control hierarchies were designed around steady-state assumptions, and in that context they deliver real value. Regulatory PID control provides stability. Model predictive control handles interacting variables. Real-time optimization (RTO) periodically computes optimal setpoints.
Each layer does its job. The optimization layer, though, was built to find the best operating point when constraint boundaries hold still, and in vacuum service they rarely do.
RTO uses steady-state models and finds solutions where constraint boundaries intersect. That works when the boundaries are current. In practice they drift as feed composition changes, exchangers foul and ambient conditions shift. In vacuum service, where condenser performance and ejector efficiency decline between turnarounds, the constraint picture can change meaningfully within weeks. The gap is less a flaw in RTO than a mismatch between how fast the plant changes and how often the model is updated.
The cost of conservative targets
Operations teams compensate by building conservatism into their limits. They hold overhead pressure targets higher than the vacuum system could currently support, keep flash zone temperatures further below the thermal limit than the current feed requires, or keep stripping steam rates up when conditions do not call for it.
Those choices are reasonable when the true constraint picture is unclear. Each increment of unneeded conservatism still adds energy cost and leaves valuable material in the residue. For many vacuum columns the core problem is coordination: too many interacting variables move at once for manual coordination or linear models to follow, and the economic optimum moves faster than an engineer can recalculate it by hand.
How a learned model coordinates the interacting variables
A model learned from the unit's own operating history and laboratory results, guided by first principles, predicts how a move in one variable affects the rest of the column. That matters in vacuum service, where the important trade-offs are rarely linear, rarely constant for long and rarely captured well by steady-state assumptions.
Constraint management is where the difference shows. LP-based systems treat constraints as fixed limits, usually set during commissioning or a turnaround and updated infrequently. A model trained on plant data learns how those boundaries behave under current conditions, including the nonlinear interactions a linear model cannot represent.
The model does not take over operator judgment. It gives operators and engineers a current view of how much room the process has, and setpoints based on learned constraint boundaries can recover margin that fixed targets leave behind.
From advisory mode to closed loop
Work on a vacuum unit usually starts in advisory mode. The model recommends setpoint moves, and operators weigh them against their own experience. Teams use the recommendations to test what-if moves when throughput, quality, energy and equipment limits pull in different directions. The same view shows where the column is losing flexibility over time as exchangers foul or feed quality shifts.
When the model suggests a move an experienced operator would not have considered, the conversation that follows often exposes an opportunity neither side would have found alone. That is how trust is built, and it is what moves a plant toward closed-loop AI adoption. In closed loop, the controller writes setpoints to the APC or distributed control system (DCS) within limits the operators define, and operators keep override authority. ARC Advisory Group describes this arrangement as human on the loop: the AI calculates and executes, and the operator moves from manual controller to high-level supervisor.
Operator judgment still matters. Conditions outside the historical training data, or equipment states the model has not seen, need experienced operators to recognize them and respond.
A shared view across functions
When maintenance, operations, planning and engineering work from the same current picture of column behavior, turnaround and constraint decisions improve. Maintenance can see where fouling is tightening constraints. Planning can compare its targets with what the column can actually do. Operations can show why a conservative move is protecting yield. That shared view closes the gap between outdated yield assumptions and the real operating boundaries, and it leads to more realistic production efficiency targets across functions.
Frequently asked questions
What does advisory mode involve on a vacuum column?
The model recommends setpoint adjustments and operators decide whether to act. Teams can test what-if moves when throughput, quality, energy and equipment limits conflict, and every shift starts from the same view of the column before any move is made. Advisory mode is where the operating team checks the model against its own experience before the controller moves to closed loop.
Does this work alongside existing APC?
Yes. The DCS keeps regulatory control and APC keeps multivariable control. The learned model works on top of both and updates the targets they work toward as feed, fouling and ambient conditions move the constraint boundaries.
How the Imubit Platform applies
In refining, Imubit applies Map, Model and Control to the vacuum unit. Map traces lost gas oil recovery and excess conservatism to the recurring decisions behind them. Model uses Deep Learning Process Control® (DLPC®) to learn a nonlinear dynamic model of the column from its own history, guided by first principles. Control takes pressure, temperature and steam decisions to closed loop, writing to the existing APC or DCS, inside limits set by plant engineers.
