Decoding lithium's subtle phase stability with a machine learning force field

Abstract

Understanding the phase stability of elemental lithium (Li) is crucial for optimizing its performance in lithium-metal battery anodes, yet this seemingly simple metal exhibits complex polymorphism that requires proper accounting for quantum and anharmonic effects to capture the subtleties in its flat energy landscape. Here we address this challenge by developing an accurate graph neural network-based machine learning force field and performing efficient self-consistent phonon calculations for bcc-, fcc-, and 9R-Li under near-ambient conditions, incorporating quantum, phonon renormalization and thermal expansion effects. Our results reveal the important role of anharmonicity in determining Li's thermodynamic properties. The free energy differences between these phases, particularly fcc- and 9R-Li are found to be only a few meV per atom, explaining the experimental challenges in obtaining phase-pure samples and suggesting a propensity for stacking faults and related defect formation. fcc-Li is confirmed as the ground state at zero temperature and pressure, and the predicted bcc-fcc phase boundary qualitatively matches experimental phase transition lines, despite overestimation of the transition temperature and pressure slope. These findings provide crucial insights into Li's complex polymorphism and establish an effective computational approach for large-scale atomistic simulations of Li in more realistic settings for practical energy storage applications.