MRS Conference (2026)
Hongjun Yang, Changyoung Park, Sungmoon Ko, Sehui Han
Abstract
Geometry-aware local message passing and global attention are combined to improve the balance between accuracy, efficiency, and simulation stability for MLMD. The architecture uses explicit geometric edge encoding for local interaction physics and a lightweight global branch for structure-level context. By avoiding heavier high-order equivariant pipelines, the model targets a practical regime with competitive accuracy and much higher simulation throughput. The model predicts energy first and derives forces and stress through differentiation, supporting physically consistent outputs for stable MD rollout. The main goal is not only low static error, but also a usable and scalable MLIP for large-system simulation.