Anorthite-NAD: Nonequilibrium Atomistic Dynamics under Controlled Deformation

Abstract

Anorthite-NAD is a molecular-dynamics dataset for studying learned nonequilibrium atomistic dynamics under controlled deformation. It contains 36 independent LAMMPS trajectories and 18,000 graph-state transitions of crystalline anorthite (CaAl₂Si₂O₈) at 300 K, spanning isotropic and axis-specific deformation at final strain magnitudes of 0.5%, 1.0%, and 1.5%. Each trajectory records atomic species, positions, velocities, forces, simulation-cell geometry, and loading information, enabling construction of graph-based state transitions for machine-learning models. The dataset uses trajectory-level train, validation, and test partitions based on independent thermal initializations to prevent adjacent MD frames from being split across partitions. Anorthite-NAD is the initial controlled benchmark for a broader effort to develop transferable equivariant graph neural networks for nonequilibrium atomistic dynamics. It is intended for experiments involving state representation, interatomic message passing, geometric equivariance, graph-perception architecture, short-horizon transition prediction, and future state-adaptive equivariant models. The current release represents a single-material, controlled-deformation benchmark and should not be interpreted as a shock, impact, or general-purpose multi-material dynamics dataset. Future extensions are planned to introduce additional materials, thermodynamic conditions, loading modes, out-of-distribution tests, and material–material interaction trajectories.