Each point pairs a generated packing with a DFT relaxation. DFT energies are relative to the lowest deposited value. Published data ↗
01
The question
A molecule can pack into many different crystals. Finding promising solid forms requires both a broad search and accurate energy estimates. The expensive step is relaxing candidate structures with density functional theory (DFT). Which candidates should receive that calculation next?
02
The approach
Our model learns the energy correction between force-field minima and DFT-relaxed structures. SOAP representations and a committee of ridge-regression models provide predictions with uncertainty. An active-learning loop selects candidates, calculates their DFT energies, and updates the model until a chosen convergence threshold is reached.
03
What we found
In the fentanyl benchmark, the paper reports 34% savings in reranking computational cost and recovery of 173 of the 174 structures in the target 1 kcal/mol window. A further study examined 17 pharmaceutical compounds using preselected structure pools.
04
Where the result applies
The method reranks an existing pool; it cannot recover a packing that was never generated. The study considers zero-kelvin energies, without finite-temperature sampling. The 17-compound exercise uses reduced pools and is not a direct comparison with a complete GRACE search.