Machine learning methods for finite-temperature full-quantum simulations: Predictive modelling of condensed phases and interfaces

Simulating complex materials, particularly interfacial and confined systems, is challenging due to the interplay of quantum mechanics in both electrons and nuclei. Here, I will present recent progress in developing efficient and accurate methods to incorporate these quantum effects. I will highlight highly data-efficient approaches for constructing machine learning potentials using only a few tens of reference structures and equivariant models of electronic properties, such as polarization and polarizability tensors, for spectroscopy. I will finally discuss the development of secondary potentials that capture quantum nuclear corrections to Born-Oppenheimer surfaces, enabling approximate quantum dynamics at a classical computational cost. These advancements make full quantum simulations computationally feasible.

As an application of these methods, I will explore the phase behaviour of water confined within nanometer-sized cavities—a model system with implications for water treatment and energy technologies. Using a predictive approach that integrates electronic structure theory, machine learning, and statistical sampling, we investigate a water monolayer confined within graphene-like channels. Our findings reveal that monolayer water exhibits rich phase behaviour and is highly sensitive to van der Waals pressure. Beyond ice phases that break traditional ice rules, we predict a superionic phase under milder conditions than those required in bulk, with electrical conductivity surpassing that of many battery materials. Notably, quantum nuclear motion significantly lowers the onset of superionic behaviour. Our work demonstrates that nanoconfinement offers a promising avenue for exploiting superionic water at near-ambient conditions.