Cause of accelerated rotational dynamics of cubes in neural-net assisted MD at low temperatures

Determine the underlying reason why, at low temperatures (T = 0.5 and 0.3 ε/kB), cubes in neural-net assisted molecular dynamics simulations—where forces and torques are computed as gradients of a trained energy neural network mapping center-of-mass separation and relative orientation to interaction energy—exhibit slightly faster rotation (larger mean squared rotation) than cubes in traditional composite rigid-body molecular dynamics simulations using explicit bead–bead interactions.

Background

The paper introduces a neural-network-assisted approach to simulate anisotropic particles by predicting interaction energies, forces, and torques directly from center-of-mass distances and relative orientations, bypassing costly bead–bead distance calculations in traditional rigid-body MD. Validation is performed by comparing structural and dynamic properties of cubes and cylinders against HOOMD-blue simulations with explicit composite beads.

While most structural metrics (e.g., pair correlation functions) and dynamic measures (mean squared displacement) match well between methods, the authors observe a discrepancy in rotational dynamics for cubes at lower temperatures: the neural-net assisted simulations yield slightly faster rotations than traditional simulations. The authors explicitly state that the origin of this discrepancy is unclear and call for further investigation, making this an explicit unresolved question.