Gabriel Krenzer1,Johan Klarbring1,2,Kasper Tolborg1,Chang-Eun Kim3,Hugo Rossignol4,Andrew McCluskey5,Benjamin Morgan6,Aron Walsh1,7
Imperial College London1,Linköping University2,Lawrence Livermore National Laboratory3,Trinity College Dublin, The University of Dublin4,European Spallation Source ERIC5,University of Bath6,Ewha Womans University7
Gabriel Krenzer1,Johan Klarbring1,2,Kasper Tolborg1,Chang-Eun Kim3,Hugo Rossignol4,Andrew McCluskey5,Benjamin Morgan6,Aron Walsh1,7
Imperial College London1,Linköping University2,Lawrence Livermore National Laboratory3,Trinity College Dublin, The University of Dublin4,European Spallation Source ERIC5,University of Bath6,Ewha Womans University7
Understanding ion transport in the superionic regime and the underlying physics of the superionic phase transition remains a long-standing challenge for computational chemistry and materials science. This presents a barrier to the development and optimisation of superionic solid electrolytes.<br/>There has also been much attention given recently to material descriptors that explore the connection between lattice dynamics and ion diffusion. Moreover, in some materials the superionic transition is accompanied by significant changes in vibrational spectra.<br/>We use <i>ab initio</i> lattice dynamics calculations to demonstrate that harmonic and quasi-harmonic descriptions of the phonons in lithium nitride show no change in features across the superionic transition. The anharmonic model, however, exhibits a breakdown for all modes. The implications for developing anharmonic lattice-dynamics-based descriptors to accelerate the discovery of superionic conductors are discussed.<br/>To further explore the superionic phase transition of lithium nitride and accurately simulate diffusion above and below the superionic transition temperature, T<sub>s</sub>, we take advantage of the recent developments in the field of Machine Learning Force Fields (MLFF). We train a Gaussian Approximation (GAP)-style MLFF using on-the-fly sampling of reference configurations. We demonstrate that our MLFF trained for lithium nitride (T<sub>s</sub>=678K) offers near-<i>ab initio</i> accuracy at a significantly reduced computational cost compared to direct <i>ab initio</i> molecular dynamics simulations. Crucially, the MLFF allows us to accurately simulate the long-timescale diffusive behaviour of lithium nitride for temperatures as low as 400K, where ion transport is slow.<br/>Using our trained MLFF, we characterise lithium nitride above and below the superionic transition temperature by calculating the heat capacity, Li<sup>+</sup> ion self-diffusion coefficient, and Li defect concentrations as functions of temperature. We show that both the Li<sup>+</sup> self-diffusion coefficient and Li vacancy concentration follow distinct Arrhenius relationships in the normal and superionic regimes. The activation energies for self-diffusion and Li vacancy formation decrease by a similar proportion across the superionic phase transition. This result suggests that the superionic transition may be driven by a change in defect formation behaviour, rather than changes in Li transport mechanism. This insight may hold implications for other superionic materials.<br/>We address a question at the heart of materials chemistry for over a century –why do some materials exhibit a superionic phase transition? This is linked to the design of solid-state batteries to address some of society’s biggest problems around energy storage. The application of machine learning methods and the resulting novel understanding of the superionic phase transition in terms of both lattice dynamics descriptors and defect formation behaviour are of interest to experimental and theoretical researchers within materials chemistry, from both academic and industrial backgrounds. This research is deeply relevant to the wide audience of the symposium on emerging challenges and opportunities in materials by design.