Energy and carbon optimisation on the crude distillation unit.

Margin in hand costs energy and carbon on every barrel.

The crude distillation unit is the largest single energy user on a refinery and one of its leading scope one carbon sources, because most of the heat a barrel needs is raised by burning fuel in the fired heater. Firing rate and excess air, pump around heat recovery, preflash and feed preheat, stripping steam, column pressure and cut points all interact, so a move that protects one product specification can quietly waste fuel elsewhere. To hold every specification safely across changing crude slates, operators run to conservative fixed set points with margin in hand. That margin costs energy and carbon on every barrel, and routine historian trends do not show where it is being lost or how much could be recovered without breaching a specification.

A Pareto set of policies, not a single answer.

EntroMetrix builds a physics informed machine learning model of the unit's energy and carbon from the plant's own historian data, combining first principles heat and mass balances with learned corrections, then takes an exergy view to locate where useful energy is destroyed rather than where it merely flows. A reinforcement learning agent works against that model and continuously trims the operating levers, furnace excess air and firing, pump around duties, preflash and stripping steam, pressure and cut points, to lower energy per barrel and carbon while every product specification is held and the unit stays inside its safe operating window. Rather than a single answer it returns a Pareto set of policies that trade energy and carbon against throughput and product value, and the plant chooses the point it wants.

Built from the historian, validated to within three percent.

The model is built from data the plant already holds. Historian records of temperatures, flows, duties and fuel across the crude unit are enough to build an energy and carbon picture of the process, and a second law, or exergy, view then shows where that energy is destroyed rather than merely where it flows. The operating limits and product specifications are set with the plant's own engineers before any recommendation is made.

On a crude distillation unit modelled and validated to within about three percent of measured temperatures, the optimised policy lowered energy per barrel by about five percent and carbon per barrel by a similar margin, chiefly by trimming fired heater firing and excess air and by tightening stripping steam, while holding every product specification. Most of the saving came from a small number of levers, so the value is reached with instruments the plant already has.

Retrospective and offline first, before any set point reaches the board.

The work is retrospective and offline first. The agent is trained against a fast physics informed surrogate of the unit and tested against operating history before any set point is suggested, and its policy is constrained to product specifications and the safe operating window, so proposals stay operable. Nothing connects to the control system in the proof phase, and the agent's own critic flags when a crude slate change has moved the unit and a refit is due.