Water in Crystals: A Database for ML and a Knowledge Base for Vibrational Prediction

Abstract

Hydrate crystals are excellent reference systems to learn about aqueous systems. We have created a database of density functional theory (DFT)-optimized (optPBE-vdW) structures and vibrational frequencies for 101 crystalline hydrate and hydroxide bulk systems and over 300 unique oscillators and use it to explore and discuss the tradeoff between prediction accuracy and insight. Starting from a machine-learning geometrical descriptor, we gradually include more physics/chemistry flavor in the descriptor and examine how the frequency prediction power varies. The most accurate models are the machine-learned model (of modest insight) and a physically motivated model containing the electric field and field gradient. Furthermore, detailed comparisons with experimental correlations show that, where available data exists, our DFT results largely overlap with the experiment. A small blind-test showed that our machine-learned (ML) descriptor model can be used to predict experimental vibrational frequencies based only on the experimental structures and our best-regressed model, with encouraging results.