Put a well-tuned MPC on an extraction tower and it will do exactly what you asked of it. Do the same on purification, and again in the sugar house. Each controller will hold its process better than anyone can hold it by hand on a night shift in week 9 of a campaign.

And the factory as a whole can still be running below what it is capable of.

That is not a contradiction. It is what happens when every stage optimises for itself.

Why Can a Factory With Good MPC Still Underperform?

The extraction controller runs in 1 of 2 modes. The first maximises sugar recovery, pushing extraction as far as the equipment limits allow. The second minimises non-sugar extraction, which lowers the load arriving at purification. Both modes are correct, on different days.

So who chooses? Today a person does, when anyone does at all, and more often the mode sits where somebody left it in January while the beet changes underneath it.

The same pattern repeats down the line. Run the evaporators hard for steam economy and you can starve the pans of the thick juice they need an hour later. Push recovery in extraction on a day when purification is already burning extra lime to keep the juice on spec, and you have moved cost from 1 station to the next rather than removed it.

The information that would settle these questions already exists in the plant. Extraction knows what quality of juice it is about to produce, purification knows what that juice is costing it in lime, and in the sugar house somebody can tell you which pans come free in 3 hours. None of it crosses the boundary between the controllers, because there was never a path for it to cross.

What Does the Coordination Layer Actually Do?

The coordination layer is a control layer above the process MPCs. It does not replace them, and it is not one giant model of the entire factory.

It takes the production targets: tonnes of white sugar, the quality grade to hit, the steam budget, the planned stops, and translates them into objectives and constraints for each process MPC. It reads back what each unit can actually deliver in the state it is in right now. Then it does the arithmetic nobody can do reliably in their head: which combination of local objectives produces the best factory result, updated as the day moves.

The shape of it will be familiar to anyone who has worked with a DCS: it is a control cascade, 1 level further up. The top controller sets setpoint objectives for the controllers below it, and those ripple down. MPC sets targets for the PLC loops without taking over the loops, and the coordination layer sets objectives for the MPCs without taking over the MPCs. Each level works at the timescale it is good at: a PID loop in milliseconds, a process MPC every few seconds to a minute, the coordination layer over hours.

The cascade also answers the first question every plant manager asks. Any single controller can be disconnected and the level below it keeps working. Take the coordination layer out and you are back to the process MPCs you had before.

How Do Extraction, Purification, Crystallisation, and Scheduling Connect?

In extraction, the coordination layer picks the objective mode instead of leaving it to whoever was in the control room. When the sugar house is the bottleneck, there is no point pushing recovery upstream; the better move is to send purification a cleaner juice.

The purification MPC simultaneously coordinates multiple chemical dosing, recycle, and temperature variables. It hits the juice quality it was asked for, and it has no way of knowing that the load arriving from extraction will change in 40 minutes, unless something tells it.

In crystallisation, the controller stabilises supersaturation using process models together with vision and near-infrared measurements. Its constraint is rarely the chemistry. It is the thick juice available, the pan that is still occupied, or the level in the strike receiver.

Scheduling is the part people underestimate. It decides when a strike starts, which crystalliser and which centrifuge it occupies, and for how long, which means a decision taken now blocks equipment for hours. The scheduler uses the current and predicted state of tanks, crystallisers, and centrifuges. It is where the factory's time dimension becomes explicit.

The mechanism that ties these together is simpler than it sounds: the coordination layer uses forward-looking information from each process area, so downstream controllers can prepare before the changes arrive. Extraction publishes where it expects to be in an hour. Purification can then treat that as a forecast disturbance it plans around, instead of a surprise it reacts to. Every one of these controllers is already predicting the future for its own use. Coordination mostly means letting the others read it.

What Changes When the Factory Has a Single Objective?

The trade-offs stop being implicit.

Right now they are being made, constantly, by whoever set a mode or a setpoint last, without anyone writing down what was traded for what. With a coordination layer, the trade-off is made once, at the top, against a stated target, and every unit works to the same definition of better.

The system will not invent a preference for you. If production, process engineering, and management do not agree on whether today is a throughput day or an energy day, the coordination layer has nothing to optimise towards. We wrote in our post on the shift from PID to MPC that the hard part of these projects has never been the mathematics. A coordination layer makes that unavoidable, which we consider a feature.

That agreement also has to take a concrete form, and the form is a single plant-level KPI system rather than 1 set of numbers per department. This is the unglamorous prerequisite, and usually the piece a factory has not built.

Where Is This Today?

The process MPCs are production reality. MPC runs on 6 extraction towers across 4 factories, and during the last campaign 1 of those controllers ran for 65 consecutive days without manual intervention.

The coordination layer above them is not at that stage. First integration tests are running, and the scheduling work runs alongside the existing system rather than in place of it, so that we can compare the 2 against real data instead of against a simulation. The architecture is defined and the first integration tests are underway.

Extraction, purification, and whole-chain scheduling come first, and the layer that coordinates them comes last. Coordinating controllers that are not yet finished would mean optimising a factory we cannot yet fully operate. A campaign is 100 to 150 days of continuous running, and every stopped hour is beet you cannot get back, so nobody sensible switches the factory's decision logic over in week 6 to see what happens.

An emerging capability with a settled architecture, not a finished product.

What Will Full Factory Coordination Look Like?

Production planning states what the factory should deliver this week. The factory works out how, across every stage, and keeps working it out as the beet and the equipment change under it. Operators supervise the factory rather than a set of units, and the questions that reach them are the ones a model was never told about: a fouled heat exchanger, a bad delivery of beet.

The part that matters commercially is that this is portable. A coordination layer built on process models rather than on 1 plant's wiring can move to the next factory, which is the difference between a project and a platform.

Want to See the Deployment Results?

The deployment methodology and results from our extraction and crystallisation projects are in the Sucrosphere white papers.

Frequently Asked Questions