Scholarlink

<p>One universal potential for all-purpose atomic simulations has been pursued for decades, but has faced extreme challenges in both reaching high representation capability of the model and constructing comprehensive potential energy surface (PES) data across the periodic table. Here we present a low-cost and high-accuracy machine learning (ML) model, namely high-order pair-reduced neural network (HPNN), which adopts a hierarchical angular interaction scheme with reduced pair dimension, enabling the incorporation of spherical harmonics up to l = 6 at a low cost. Using the architecture, we demonstrate the capability to train a generalized global neural network potential (GG-NN) using a comprehensive dataset of 5.84 million global PES configurations covering 83 elements in the periodic table collected from LASP project in the past decade, which reaches the root-mean-square errors of 7.3 meV/atom for energy and 0.16 eV/Å for force. By benchmarking with mainstream ML models on representative global PESs, e.g., Ti–O and C–H–O, we show the high performance of specially-trained HPNN and GG-NN both in the inference speed and in the accuracy. Our results pave the way to exploit the generalized global potential for large-scale atomic simulations to accelerate molecules and material prediction via global PES exploration.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

High-order pair-reduced neural network architecture for global potential energy surface exploration across the periodic table

  • 期刊论文

  • 发表日期：

    2025年11月11日

查看全文
  • Zheng-Xin Yang,
  • Xin-Tian Xie,
  • Zhen-Xiong Wang,
  • Dong-Xiao Chen,
  • Zi-Xing Guo,
  • Jia-Jie Du,
  • Qi-Ming Liang,
  • Qian-Yu Liu,
  • Cheng Shang,
  • Zhi-Pan Liu

摘要

One universal potential for all-purpose atomic simulations has been pursued for decades, but has faced extreme challenges in both reaching high representation capability of the model and constructing comprehensive potential energy surface (PES) data across the periodic table. Here we present a low-cost and high-accuracy machine learning (ML) model, namely high-order pair-reduced neural network (HPNN), which adopts a hierarchical angular interaction scheme with reduced pair dimension, enabling the incorporation of spherical harmonics up to l = 6 at a low cost. Using the architecture, we demonstrate the capability to train a generalized global neural network potential (GG-NN) using a comprehensive dataset of 5.84 million global PES configurations covering 83 elements in the periodic table collected from LASP project in the past decade, which reaches the root-mean-square errors of 7.3 meV/atom for energy and 0.16 eV/Å for force. By benchmarking with mainstream ML models on representative global PESs, e.g., Ti–O and C–H–O, we show the high performance of specially-trained HPNN and GG-NN both in the inference speed and in the accuracy. Our results pave the way to exploit the generalized global potential for large-scale atomic simulations to accelerate molecules and material prediction via global PES exploration.