Machine-learning ice spectra: from 1 to 256 features

The study explores how well machine learning can predict spectroscopic properties of ice (OH vibrational frequencies and ¹H chemical shifts) using a range of structural descriptors (and some regressors). A large theoretical dataset (55 ice polymorphs, 1000 DFT data points both for the vibrations and for the NMR shifts) and a smaller cross-validation set are employed. The Message Passing Atomic Cluster Expansion (MACE) model performs best, with high accuracy (root-mean square deviation, RMSD, of 0.06 ppm for chemical shifts and ~10 cm⁻¹ for vibrational frequencies). Simpler descriptors like ACSF and SOAP, when paired with suitable regressors, nearly match MACE’s performance. At the other end of the complexity scale, it is found that using the simplest possible physics-based descriptor of the environment (a single H-bond distance) yields RMSD values three times as large for the vibrations and seven times as large for the proton chemical shift compared to the MACE model. Depending on the context, those RMSD values may still be considered modest and useful, considering the gain is simplicity and transparency.