Three sentences describe how Lavoisier builds every prediction:
First principles. Then ML. Always uncertainty.
First principles
Start with thermodynamics, mixing rules, and validated property models where they apply. Ideal systems stay physically plausible even when data is thin. Chemists already trust this layer — it is the language of the plant, not a neural net weight dump.
Then ML
Real formulations are non-ideal. Synergies, impurities, and process history break pure theory. Machine learning learns the residual: what the baseline misses, within a domain you can measure. That is how hybrids stay useful without inventing non-physical viscosity curves.
Always uncertainty
A point prediction without confidence is a liability in a design review. Show intervals, extrapolation flags, and when not to trust the number. Uncertainty is not a disclaimer — it is part of the answer.
Why this order matters
Flip the stack and you get black-box fluency with no mechanism. Skip uncertainty and you get false precision. Keep the order and chemists can challenge the answer — which is the only way AI survives next to capital decisions and plant SOP.
