Hybrid computational approaches for accurate crystalline density prediction of cyclo-pentazolate salts

Abstract

In this paper, we present three computational approaches for accurately predicting the crystalline densities of cyclo-pentazolate salts based on three level chemical descriptions, encompassing both metal (Li, Na, K, Mg, Ba, Al, Cu, Ag, Zn, Co, Fe, Mn, and Pb) and organic cations, based on the most comprehensive list of experimentally available crystals (69 examples). The level 1 description involves force-field-based crystal structure prediction. Level 2 employs semi-empirical PM7 calculations of vacuum-isolated crystal bases. Level 3 focuses on empirical formula analysis. Each approach generates a unique set of descriptors used to develop corresponding empirical models through machine learning techniques. The results demonstrate excellent predictive performance for cyclo-pentazolate salts across a wide range of compositions, from simple binary ionic pairs (Cat⁺An⁻) to complex multi-ionic structures containing neutral molecular additives. The mean absolute percentage errors (MAPEs) across all models range from 3% to 5%. Prediction at level 3 requires no quantum-chemical or molecular-mechanics calculations, making it applicable to high throughput predictions of any potential cyclo-pentazolate salts without significant structural voids, such as nanocages or nanopores. For novel salts with unusual ionic and/or molecular components, validation using levels 1 and 2, which are based on robust independent methodologies, is recommended.