Angular and Shell-Aware Deep Potential Energy Model for Molecular Dynamics

Yuan Pu, Jiayang Yao, Haoxin Wang, Binwu Zhu, Yuyang Ye, Zhuolun He, Yian Wang, Bei Yu

Keywords: Machine Learning Potentials (MLP); Potential Energy Surface (PES); Molecular Dynamics (MD); Deep Potential (DP); Quantum Mechanics (QM) Accuracy

TL;DR: This paper introduces ASDP, a novel deep potential that integrates a shell-aware angular bias into an attention mechanism for highly accurate potential energy surface modeling.

Abstract:

Angular information, especially involving the first and second coordination shells, is critical for accurately describing the potential energy surface (PES) in molecular systems.
However, existing machine learning PES models either neglect this information or indiscriminately process it from all neighbors, blurring the critical contributions of distinct shells and compromising their predictive accuracy.
In this work, we propose the Angular and Shell-Aware Deep Potential (ASDP), a novel architecture designed to overcome this limitation.
Based on the DPA-1 attention mechanism, ASDP integrates a specialized encoding module that selectively processes angular information confined within the first two coordination shells.
This shell-aware approach allows for a more physically meaningful representation of the local atomic environment.
Experimental results show that by capturing crucial shell-specific angular dependencies, ASDP represents the PES of various molecular systems with the \textit{ab initio} quantum mechanics (QM) accuracy,
outperforming many existing methods and offering a new direction for creating highly accurate and robust machine learning potentials.
Our code can be found in \url{https://anonymous.4open.science/r/ASDP-ICLR-code}.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 10348