Projects
Selected work, each with its own page. Click a card to read more.
Rem3Di
2026Learning smooth, chiral 3D molecular descriptors from atomistic foundation models
Rem3Di repurposes latent features from atomistic foundation models as transferable molecular descriptors. It pools a potential's per-atom features into one fixed-length descriptor that varies smoothly with 3D structure, and adds pseudoscalar channels so the representation can tell enantiomers apart. It matches or beats published baselines on drug-property benchmarks without any 2D fingerprints.
Accurate energy barriers for catalytic reaction pathways: an automatic training protocol for machine learning force fields
An automatic, active-learning training protocol that gets machine-learned force fields to DFT-quality energy barriers for catalytic reaction pathways. On CO₂ hydrogenation to methanol over indium oxide it lands within 0.05 eV of DFT — and finds an alternative path that cuts the previously established rate-limiting barrier by 40%.
MACE-MP-0
2024A foundation model for atomistic materials chemistry
A single general-purpose force field, trained on a public dataset of ~150k inorganic crystals, that runs stable molecular dynamics across molecules and materials — solids, liquids, gases, reactions, interfaces, even a small protein. It works out of the box as a starting point for any atomistic system, and fine-tunes to ab initio accuracy on a handful of points.