Abstract
<jats:p>General-purpose AI applied to chemical engineering routinely violates conservation laws and extrapolates poorly. This work presents domain-informed AI as a framework organizing how chemical engineering knowledge enters an AI architecture. Most techniques reduce to two strategies: architectural modification, altering network structure through constrained outputs, inductive-bias neurons, and hybrid differential equations; and input engineering, structuring inputs through graph representations, transfer learning, and multimodal fusion. Cutting across both is a distinction from physicsinformed learning between constraints guaranteed by construction and constraints only encouraged by a loss term, argued here to be what determines when a model can be trusted. Domain-informed AI is proposed as a superset of informed, physics-informed, and scientific machine learning and of hybrid modeling. That claim is tested: at the systems level neither strategy applies, since flowsheet validity is discrete rather than differentiable, and knowledge enters instead by external delegation to a solver or validator. Four tiers are treated with deliberately uneven demonstration. On a CSTR study a selective hybrid reaches a trajectory error about 5× lower than a Neural ODE in distribution and about 140× lower outside the training envelope, while the learned reaction rate is only weakly identifiable. On molecular property prediction a graph neural network reaches R2 = 0.911± 0.015 on ESOL, and transfer to the data-scarce FreeSolv target lowers error by about 14% at 10%</jats:p>