Research

A committee-of-experts approach combining specialised machine-learning potentials for predictions of atomic energies and forces.

An investigation of extrapolation in high-dimensional machine-learning potentials, using the geometry and probability density of the representation space.

Automatic selection of atomic fingerprints and reference configurations to reduce the dimensionality and training requirements of machine-learning potentials.

A comparison of molecular descriptors for predicting solubility in lipid excipients and interpreting solute–lipid interactions.

A machine-learning framework for predicting the electronic density of states through local, atom-centred contributions, studied in condensed-matter systems.

A Bayesian framework combining predicted NMR chemical shifts with uncertainty estimates to assess candidate molecular crystal structures.

A generalised convex-hull framework using structural descriptors to identify candidate materials and explore their stability under different constraints.

A community benchmark comparing crystal structure ranking methods across pharmaceutical targets, including dispersion-corrected DFT and machine-learning approaches.

A community benchmark of crystal structure generation methods, examining the recovery of experimental structures across pharmaceutical targets.

An active-learning workflow that uses predicted energies and uncertainties to select crystal structures for DFT relaxation. The fentanyl benchmark reports 34% savings in reranking computational cost; a separate reduced-pool study covers 17 pharmaceutical compounds.

SMolNet learns embeddings of experimental XRPD patterns for pairwise solid-form comparison, evaluated on compounds withheld from training.

APPA combines scientific literature, experimental databases and machine-learning predictors to support evidence gathering and preformulation planning for drug candidates.