Least cost gasoline blends that stay on spec.

Blending is nonlinear, so grades are over delivered.

Gasoline blending is nonlinear. Octane and vapour pressure do not blend linearly by volume, so a finished property cannot be read off from a simple weighted average of the components. Component properties also drift as upstream units move and feedstocks change, and analysers report only every half hour. Because the true blend behaviour is hard to predict, recipes are set conservatively and grades are over delivered on octane and on vapour pressure. That margin is value given away on every blend. When a recipe still lands off spec, as when a component is switched, the blend has to be corrected or reblended, which consumes capacity and raises cost further.

Property models of the blend and an agent that proposes recipes.

EntroMetrix builds physics informed property models of the nonlinear blend behaviour, combining simulations and mechanistic blend models with the plant's own measurements and analyser records to capture how octane and the other quality properties respond to each component. A reinforcement learning agent then proposes component recipes online, moving the recipe in small mass balanced steps that hold every grade specification while lowering cost and shrinking giveaway. Because it learns the blend mechanism from data, a component switch or a shift in component quality no longer forces an octane overshoot or a spec breach, and the specifications are held as constraints by a safe reinforcement learning layer that keeps the policy inside grade, tank and component limits. The system is retrospective and offline first, scored against the plant's past blends before any live use.

Built from analyser records, demonstrated on nonlinear blending.

The models are built from data the plant already holds. Component and product analyser records, tank movements and the blend history are enough to learn how octane and the other properties respond to each component, and the grade specifications together with the tank and line limits are set with the blending team. Assumptions are documented and reviewed before any recipe is proposed.

The method has been demonstrated on nonlinear gasoline blending. A property model of the kind used here predicts octane from the component recipe to within a few tenths of a research octane number, and against that model the agent recomputes a full on spec recipe in seconds. Where a conventional linear recipe overshoots octane and breaches a component limit after a feed change, the learned policy holds every specification through the same change, which is where the giveaway and reblends come out.

Scored against past blends, proposed to the blender.

Validation is retrospective and blind. Recipes are scored against the plant's own past blends before any live use, and the policy is constrained to stay inside grade, tank and component limits, so proposed moves remain within ranges the blender already accepts. The system runs offline, isolated from control systems, and proposes recipes for the blender to approve rather than acting on the plant, with the cost, giveaway and on spec position of each recipe shown.