Now in design-partner pilots · Battery, polymer, pharma
SOLVOLYTE
From years of bench screening to a shortlist in weeks. Physics-constrained AI for the chemistry of liquid mixtures.
100k
training datapoints
0.83
R² out-of-distribution
10⁵
Na-ion screened
From model to shortlist
100k
training datapoints across conductivity and viscosity
0.83
R² on held-out molecule generalization
10⁵
Na-ion electrolyte candidates screened
Trusted in pilots by
Two years of screening, in a sprint.
12–24 months on the bench.
A new electrolyte program runs months of formulation, cell builds, and cycling. Most candidates fail by week 6.
A ranked shortlist in weeks.
We take your target chemistry, screen 10⁵ candidates against ionic conductivity, viscosity, and electrochemical stability — return the top 50 with confidence intervals.
3D physics, not 2D fingerprints.
Mixtures aren't the sum of their parts. Our Set-SE(3) equivariant model learns shape-aware solvation — the part 2D GNNs throw away.
Industries we're building for.
Batteries.
10⁶⁰ candidate mixtures. Most are dead. The sweet spot is rare. We find it before the bench does.
Polymer blends.
Tg, miscibility and modulus are exquisitely sensitive to composition and processing. Blind blend screens cost months on the rheometer.
- Phase behavior depends on chain conformations, not just monomers
- Industrial blend datasets are sparse and proprietary
- Failure mode: phase separation, brittleness, processing window
- → PolySolv learns conformer-ensemble representations, fine-tuned on your in-house blend data
Drug delivery.
The same API can be inert or therapeutic depending on the co-solvent matrix. Solubility and crystallization are entangled and hard to optimize.
- Co-solvent ratios shift solubility curves non-monotonically
- Crystalline polymorphs threaten shelf life
- Excipient screening is high-dimensional and bench-bound
- → PharmaSolv predicts non-monotonic solubility curves and ranks polymorph stability before bench
Eight weeks. Fixed scope.
We engage as a paid 8-week design partnership. You bring the chemistry; we return a ranked, physics-validated shortlist plus the fine-tuned model head. IP on derived predictions stays with you.
- Week 0
Scoping call.
30 minutes. We confirm fit, target property, and what data you can share.
- Week 1–2
Data ingest + baseline.
Your private dataset lands in an isolated namespace. We fit a baseline head and report held-out R² on your data.
- Week 3–6
Domain-tuned screen.
We screen 10⁵ candidates with the tuned model and rank by your composite objective.
- Week 7–8
Validated shortlist.
Top 50 with confidence intervals, MD-grounded sanity checks, and synthesis notes for each candidate.