论文标题
磁性超结构的深度学习电子结构计算
Deep-learning electronic-structure calculation of magnetic superstructures
论文作者
论文摘要
磁性超结构（例如，磁性天际）的摘要研究对于新型材料的研究是必不可少的，但由于其强大的计算成本而瓶颈。为了解决瓶颈问题，我们开发了一种深层的神经网络方法（命名为Xdeeph），以表示密度功能理论Hamiltonian $ H_ \ Text {Dft} $作为原子和磁性结构的函数，并应用神经网络以进行有效的电子结构计算。神经网络的智能通过将有关重要位置和对称属性的先验知识纳入方法来优化。特别是，我们设计了一个神经网络体系结构，该架构完全保留了$ H_ \ text {dft} $的所有等效要求，由Euclidean和Time-vers-vers-vers-vers-eversmetries（$ e（3）\ times \ times \ times \ {i，t \} $），这对于提高方法的性能至关重要。纳米管，自旋刺激性和莫伊尔磁体的系统实验显示了Xdeeph的高精度（亚M-EV误差）和良好的可传递性，还显示了研究磁性天空敏的能力。该方法可以在磁性材料研究中找到有希望的应用，并激发了深度学习的初始方法的发展。
Ab initio study of magnetic superstructures (e.g., magnetic skyrmion) is indispensable to the research of novel materials but bottlenecked by its formidable computational cost. For solving the bottleneck problem, we develop a deep equivariant neural network method (named xDeepH) to represent density functional theory Hamiltonian $H_\text{DFT}$ as a function of atomic and magnetic structures and apply neural networks for efficient electronic structure calculation. Intelligence of neural networks is optimized by incorporating a priori knowledge about the important locality and symmetry properties into the method. Particularly, we design a neural-network architecture fully preserving all equivalent requirements on $H_\text{DFT}$ by the Euclidean and time-reversal symmetries ($E(3) \times \{I, T\}$), which is essential to improve method performance. High accuracy (sub-meV error) and good transferability of xDeepH are shown by systematic experiments on nanotube, spin-spiral, and Moiré magnets, and the capability of studying magnetic skyrmion is also demonstrated. The method could find promising applications in magnetic materials research and inspire development of deep-learning ab initio methods.
