Study of the thermodynamics and kinetic anomalies of Antimony-based phase-change materials by machine-learned molecular dynamics

Energy-intensive data centers and large-scale machine learning models, combined with the exponential growth of global data generation, are unsustainable in the long term unless more efficient data storage and computing technologies are developed. Promising next-generation devices include ultrafast non-volatile memories and brain-inspired computing units. The core of these technologies is enabled by phase-change materials, alloys of group-IV to -VI elements exhibiting unique, reversible switching properties between a resistive amorphous (glass) phase and a conducting crystal on a timescale of hundreds of nanoseconds. The design and optimization of phase-change materials are closely linked to the physics of supercooled liquids, i.e., metastable liquids below their melting points. The elucidation of the kinetics and thermodynamics of complex materials under deep supercooling is now within reach of atomistic simulations, thanks to the advent of machine-learned interatomic potentials trained on ab-initio data. In this work, we focus on antimony (Sb), a recent candidate for monoatomic phase-change materials, and Sb-rich Ge-Sb alloys. Signs of a liquid-liquid transition in Ge$_{15}$Sb$_{85}$ have been reported in femtosecond X-ray experiments supported by ab-initio simulations. Motivated by this, we investigate possible signatures of a liquid-liquid transition in elemental Sb using a machine-learning potential taken from the literature. By systematically exploring the structure of the crystals spontaneously nucleating from the supercooled liquid phase, we discover an unreported phase of Sb: a bulk A17 (black phosphorus–type) layered structure emerging upon decompression, which is the only phase to nucleate below –1 GPa. Within the supercooled liquid, we find water-like thermodynamic and structural anomalies, suggesting the presence of a liquid-liquid transition. We introduce a custom octahedral order parameter — analogous to the tetrahedral parameter for water — to describe these anomalies in terms of the emergence of a structured liquid at low temperature, and we perform a two-state model fit. Direct computation of the configurational entropy provides no evidence for a kinetic crossover, indicating that Sb is a highly fragile system. Finally, we train an machine-learning potential for Sb using an in-house architecture - \alphanes, proposed by Guidarelli Mattioli, Sciortino and Russo in 2023 -, which at ambient pressure achieves comparable accuracy to the existing potential. Building on this experience, we generate a dedicated ab initio dataset for Ge$_{15}$Sb$_{85}$, comprising nearly 200k atoms in total, on which we train a machine-learning potential via the NeuroEvolution Potential framework. Validation confirms its robustness against overfitting and its suitability for multi-phase modeling in future studies of this complex alloy.