Most formulation work still runs forward: mix, measure, adjust, repeat. Inverse formulation flips the question — describe the product you need, then search for compositions that can hit it.
What “inverse” actually means
Given targets (viscosity band, Tg, VOC limit, cost ceiling) and a catalogue of real materials, the system proposes ranked candidates. It is not inventing molecules from noise; it is searching a constrained space of purchasable components.
Why first principles matter here
Pure ML recommenders overfit to historical wins and fail when the catalogue or the specification shifts. A physics-informed stack keeps mixture baselines honest, then lets ML correct residuals where data exists — see Physics-informed vs black-box.
What chemists should demand
- Constraints as first-class inputs — REACH status, customer banned lists, processability limits.
- Confidence on every candidate — not a single “best” number without an extrapolation flag (confidence intervals).
- Inspectable rankings — why candidate A beat B, in language a formulator can challenge.
That is the FormulaIQ posture: target properties in, in-spec recipes from real materials — predictions you can defend before a drop is mixed.
