Near optimal crude and production schedules in seconds, not hours.

Exact solvers scale badly, and manual plans leave margin on the table.

Crude scheduling is one of the largest combinatorial problems a refinery solves. Planners must sequence tanker and pipeline receipts, allocate storage and charging tanks, fix blend recipes against sulphur, acidity, density and residue limits, and set the charge to each distillation unit across a horizon of one to two weeks. The schedule has to be worked again every time a tanker slips, a tank frees, a blend target moves or a unit rate changes. Exact mixed integer formulations reach true optima but scale badly, growing from under a minute at five days to tens of hours at ten days, and any disturbance forces a full recomputation. Manual planning is fast but conservative, and leaves margin on the table.

A scheduling environment on your history, and a policy that stays feasible.

EntroMetrix builds a scheduling environment on the operator's own historical data and trains a reinforcement learning policy to sequence receipts, allocate tanks, set blends and charge the units. The problem is cast as a Markov decision process: the state carries tank inventories, crude qualities and vessel arrivals, the uncertain ones held as Bayesian distributions rather than single values, the action sets receipt, transfer and charge volumes, and the reward is gross margin from a physics informed blend and yield model, net of demurrage, changeover and slop. Every action is passed through a math programming projection, so it lands on the feasible region set by tank capacities, connectivity, blend properties and unit rates before it is applied. Once trained the policy returns a schedule in seconds when a tanker slips or a unit trips, staying within a few percent of the mixed integer optimum, and surfaces a Pareto set the planner chooses from.

Built from receipt logs and tank movements, within a few percent of the optimum.

The scheduler is built from data the plant already holds. Historical receipt logs, tank movements, blend records, unit rates and the crude assay library are enough to reconstruct the operator's own decisions and to train the policy against them. The logistics and quality limits and the operating window are set with the schedulers before any run.

On a comparable crude scheduling problem the learned policy reached within a few percent of the mixed integer optimum while returning a schedule in seconds, one to two orders of magnitude faster than a full re solve, and held that margin when a receipt was delayed because it was trained against the Bayesian distribution of arrival times and could re optimise inside the horizon rather than run a stale plan. Feasibility was guaranteed throughout by projecting every action onto the region the plant can actually operate, so proposed schedules stayed inside the accepted limits rather than breaching them as a policy trained on penalties alone would.

Planner in the loop, with feasibility held by projection.

Validation is retrospective and blind. Schedules are compared against the operator's own historical schedules on held out periods before any live use, the policy is held inside logistics and process limits by a math programming projection and a constrained objective, and the scheduler runs offline for decision support, with the planner reviewing and approving the Pareto set before a schedule is issued.