Transferability and expansion of the clayNN model

We demonstrate that neural network potentials for kaolinite clay trained using forces and energies at the density functional theory level using two dispersion-corrected functionals are able to reproduce the structural properties of dickite and nacrite even when no structures for these minerals are included in the potential's training data. We also train a potential using the meta-GGA functional SCAN, which provides an extremely accurate reproduction of the structural properties of the studied kaolin minerals, and is the only potential able to discern subtle differences in polyhedral environments. We also build models for the polarisation of kaolin minerals, allowing the infrared spectra to be extracted from simulations. The ClayNN models provide an excellent description of the structural properties of the studied kaolin minerals and when combined with our polarisation models also give a reasonable description of the number of peaks in the vibrational spectra but perform poorly when considering peak positions.<br/><br/>