Formulation Studio — read a whole recipe before you pour

[rvnminers-A-and-N]

The "before you pour" bench read

A bench chemist works with formulations — many ingredients at ppm — not single molecules. The Formulation Studio plugs in the whole recipe and predicts what it will come out like, so runs get spent where they count.

What it does

• Blended note-profile — aggregates each ingredient's predicted aroma, weighted by odor impact (OAV where thresholds are loaded, else mass × volatility), into the profile the blend reads as, naming the driving ingredient under each note.
• Overpowering-component flag — "too heavy in one item," computed from the blend weights so it works in both the quantitative (OAV) and directional (volatility) bases.
• Target + gap analysis — set the notes you're aiming for; get what's short (with food-safe, GRAS-preferred carriers to add) and what's over/unrequested (what to ease back).
• Documented-hazard screen — reuses the curated precursor/process hazard rules on the full formulation.
• Honest data-gates, in the UI — directional today; calibrated intensity and suppression/synergy (a learned mixture model) unlock with the customer's odor-threshold / panel data. Missing-note carriers with no clean public source are labelled data-gated rather than faked.

Engine

• New POST /api/formulation composes the existing RF taste/aroma heads + analyze_balance (OAV) + check_mixture + palette_match.
• Note weighting, aggregation, and the gap/overpower logic are deterministic chemistry math over the ML predictions — reliable, explainable, auditable (the right posture for a safety-adjacent bench tool).

Performance

• Memoized predict_aroma and _resolve.
• Pinned the RF heads to n_jobs=1 — single-sample predict_proba was spending ~150 ms/call spawning a joblib thread pool (~3.8 s → ~1.2 s per 24-head aroma read); speeds the whole app.
• Local name→SMILES index (from master_enrichment.parquet + the suggest CSV) so library molecules resolve offline; only unknown names hit PubChem.
• Background pre-warm of the demo molecules at startup, so the starter formulas are instant.

Verified live with Playwright (three starter formulas, profile bars, overpowering flag, gap analysis, hazard path) and regression-checked /api/predict + /api/mixture.