Generalised deep-learning workflow for the prediction of hydration layers over surfaces

Atomic force microscopy (AFM) is paving the way for understanding the solid–liquid interfaces at the nanoscale. These AFM studies are complemented with molecular dynamics (MD) simulations of hydration layers over candidate surfaces for a comprehensive characterisation. We earlier proposed, in Ranawat et.al. (2021), a deep-learning (DL) network to predict hydration layers over the candidate surfaces much more rapidly than computationally-intensive MD. However, the proposed elements-as-channels based network is bound to the elements present in the training surfaces. Here, we develop a generalised descriptor of the surface to train element-agnostic networks. We demonstrate the descriptor's efficacy by predicting the hydration layers over a dolomite surface using a network trained on the calcite and magnesite surfaces. We also demonstrate the transfer-learning capability of such a descriptor by incorporating mica into the training surfaces, and predict the pyrophyllite and boehmite surfaces. Further, we propose an energy-based DL framework to gauge the possible prediction accuracy of a network on surfaces hitherto unseen. We combine these advance techniques into a generalised workflow to complement AFM studies.