AI for the chemistry of liquid mixtures.
Liquid mixtures power batteries, drug delivery, and specialty chemistry — yet most ML stops at single molecules. Solvolyte is built around a different premise: that the mixture is the unit of design, that geometry governs transport, and that a model is only as good as the bench experiment it predicts. We engage as paid 8-week design partnerships with R&D teams who need a shortlist, not a paper. Every engagement closes the loop between prediction and electrochemistry, so the model your team uses in week eight is sharper than the one it started with.
Physics-first
SE(3)-equivariant message passing enforces rotational and permutation symmetry by construction. Solvation-shell geometry, ion-pairing distances, and conformer ensembles are inputs — not learned from scratch. The model never violates the underlying physics, even out-of-distribution.
Experiment-grounded
Every shortlist is closed-loop with a synthesis and electrochemistry partner. Predictions ship with calibrated uncertainty, and the experimental return signal updates the next training round. We don't ship a number we wouldn't bet a Coulombmeter on.
Open by default
Pre-trained GeoSolv weights, the Set-SE(3) reference implementation, and the CheMixHub-derived training pipeline are released under permissive licences upon paper acceptance. Partners keep full IP on derived predictions; we keep the method open so the field moves forward.
Let's scope a pilot.
Research
[email protected]Partnerships
[email protected]Based in
Chicago, IL · Pritzker School of Molecular Engineering
Open science
Code and pre-trained GeoSolv models will be released open-source upon publication.