GXTNet: Balancing Accuracy and Cost for Molecular Dynamics

While machine learning interatomic potentials (MLIPs) have demonstrated near-DFT accuracy across a range of static benchmarks, their performance in molecular dynamics (MD) simulations also depends on factors not fully captured by static prediction errors, including rollout stability and computational cost. To address this, we propose GXTNet (Graph-X-Transformer Network), which combines GNN-based local message passing with Graph Token-Mediated Attention (GTMA), using a recurrent token for sparse, geometry-conditioned global communication. Consequently, GXTNet significantly advances the accuracy-cost Pareto front among models trained on the MPtrj dataset. Evaluated on the Dyna-Mat benchmark, it achieves a high combined MD score (CMDS) of $0.7126$. Furthermore, the model exhibits robust scaling performance, maintaining the lowest end-to-end force-evaluation costs across large-scale systems exceeding $10{,}000$ atoms.