Most solubility tools give you three numbers and false confidence. Ours is a fitted model on RDKit fragment features over a curated literature solvent set, with an error estimate and an applicability-domain flag: outside the flag, the UI says "out-of-domain" instead of inventing precision.
How to use it (5 minutes):
- Sketch or paste SMILES (Ketcher 2D sketcher built in).
- Get δD/δP/δH plus interval plus domain flag.
- Rank candidates by distance-to-target with cost attached (cost-vs-performance Pareto, not a single score).
- Export the shortlist with intervals intact — no naked point estimates.
When it fails: novel scaffolds far from the literature solvent set, strong H-bond networks, temperature dependence beyond ambient. All three are labeled, not hidden.
Pair it with: D-optimal mixture design (coordinate exchange, Scheffé linear/quadratic, D-efficiency report) so the next experiments cover the space instead of re-testing the same corner.
FAQ: Is this a replacement for measurement? No — it's PREDICTED, screening-grade triage before you run the plate. Can I bring my own data? Yes — upload measured solubilities alongside the predictions to compare and, where the model path supports it, fit on your set. Cost? Screening is free-tier; real physics (QM/MD) meters credits only.