| Abstract Scope |
Universal machine learning force fields (UMLFF) are applied to predict phase transition sequences in zirconia (ZrO₂) using molecular dynamics (MD) simulations and quasi-harmonic approximation (QHA). Two models, SevenNet (7net-0(lmax=2) and 7net-l3i5(lmax=3)) were employed in MD simulations, revealing that 7net-0 incorrectly predicted monoclinic-orthorhombic-cubic transitions, whereas 7net-l3i5, predicted experimentally observed monoclinic-tetragonal-cubic transition sequence. Additionally, 10 different UMLFF models and density functional theory (DFT) were applied via QHA to assess monoclinic-tetragonal (M-T) and monoclinic-orthorhombic (M-O) transitions. While DFT predicts the experimental M-T transition, only mace-d3 and 7net-l3i5 predict the M-T transition; others predicted M-O transitions. These findings underscore UMLFF's potential to estimate free energies of crystals close to DFT but highlights that ~4 meV/atom errors in relative ground-state energies can lead to incorrect phase stability predictions at finite temperature. High-quality training datasets as well as higher spherical harmonics are critical for UMLFF to accurately predict force, energy, and phase transitions in different materials. |