Machine-Learning Molecular Dynamics Study on the Structure and Glass Transition of Calcium Aluminosilicate Glasses

Calcium aluminosilicate (CAS) glass systems represent an important class of materials for industrial applications due to their superior thermal and mechanical properties. Although CAS glasses have been extensively studied, the structure-property relationships, particularly in the peraluminous region (Al₂O₃/CaO > 1), remain insufficiently understood. Experimental studies have identified the presence of five-coordinated aluminum (Al⁵) depending on the Al₂O₃ content; however, classical molecular dynamics simulations have struggled to accurately reproduce the aluminum coordination environment. To address this limitation, we developed a machine-learning potential tailored for CAS systems, trained on a comprehensive molecular dynamics simulation based on density functional theory data set and refined using a fine-tuning approach. Melt-quench simulations were then carried out to model CAS glass structures. The resulting structures from machine learning-based molecular dynamics (MLMD) accurately reproduced both experimental glass densities and the fractions of Al⁵, including the observed increase in Al⁵ and oxygen triclusters (TBOs) in the peraluminous region. In addition, we performed heating simulations to evaluate enthalpy changes and structural evolution as a function of temperature. Analogous to differential scanning calorimetry experiments, the glass transition temperature (<i>T</i>_g) was determined from the MLMD data. The compositional dependence of Al⁵ and TBO near the <i>T</i>_g was quantitatively analyzed, providing insights into the role of aluminum in structural rearrangements. These findings demonstrate that MLMD enables the accurate modeling of CAS glass structures and yields valuable insights into their thermal behavior. This approach offers a robust framework for understanding structure-property relationships in complex glass systems.