Coatings R&D burns time on failed experiments: gloss missed, viscosity out of window, VOC over limit, adhesion failure after humidity. The calendar cost compounds across pigment, binder, and additive choices.
What compresses the loop
- Targets first — encode the product window before mixing (inverse formulation).
- Real catalogues — only materials you can buy and register.
- Ranked candidates with confidence — spend wet work where the model is uncertain or where value is highest, not where it is already confident.
What does not compress the loop
A chatbot that suggests “try adding more dispersant” without constraints, property prediction, or uncertainty. That recreates folklore at LLM speed.
The Lavoisier bet for coatings teams
Physics-informed property prediction plus regulatory and processability constraints — so fewer failed panels, clearer trade-offs, and a paper trail for why a candidate was advanced. Shorter R&D without surrendering chemistry to a black box.
