GeoSolv
Find safer, higher-conductivity Na-ion electrolytes — at 105 screened per pilot, with calibrated confidence on every shortlist.
R²
Held-out generalization
candidates
Per Na-ion shortlist
mS/cm
Best predicted
ML for mixtures collapses on unseen chemistries.
2D screens promote candidates that fail in cells.
Fingerprint-based ML scores molecules independently. Mixtures aren't separable — pairwise solvation governs whether ions move at all. Two-formulation differences invisible in 2D ruin a cell-build at week 12.
Small composition shifts flip viscosity and stability.
Property surfaces over composition aren't smooth — they have ridges where the model needs to be sharpest. Grid search on a 6-component blend explodes; off-grid points are exactly where the gains live.
Solvation shells govern transport. We learn the shell.
Ionic conductivity and viscosity track shell geometry, not connectivity. Set-SE(3) message passing puts the 3D shell into the model — that's why GeoSolv generalises to molecules it has never seen during training.
A physics-constrained pipeline, end to end.
3D point cloud encoding
Set-SE(3) message passing
Differentiable Arrhenius readout
10⁵ candidate screening
+22 R2 over the prior best.
Arrhenius conductivity trace
top Na-ion candidateCheMixHub leave-molecules-out
R² · higher = betterEight weeks. Fixed scope. Real deliverables.
We engage as a paid 8-week design partnership — not a one-off API call, not a consulting hand-wave. You bring the chemistry; we return a ranked, physics-validated shortlist plus the model head. Prediction IP stays with you.
Your chemistry, your target.
- A SMILES list or sketch of the formulation space
- Target property + threshold (e.g. > 5 mS/cm at -10 °C)
- Optional: any in-house measurements you can share
A shortlist you can build.
- Top-50 candidates ranked by composite objective
- Calibrated confidence intervals on every prediction
- MD-grounded sanity report on the top 10
- Synthesis notes + a fine-tuned model head you keep
Your data, your namespace.
- Isolated tenant — separate buckets, IAM, KMS keys
- TLS 1.3 in transit, AES-256 at rest, hardware KMS
- Wiped on engagement end or 30 days, whichever first
- You keep IP on every derived prediction and shortlist