可转让性
哈密顿量(控制论)
半导体
电荷(物理)
密度泛函理论
物理
电子结构
材料科学
原子物理学
统计物理学
凝聚态物理
晶体缺陷
电荷密度
氧原子
能量(信号处理)
静电学
分子物理学
总能量
氧气
氮气
混合功能
计算物理学
作者
Mingjue Ni,Zhenxing Dai,Menglin Huang,Hongjun Xiang,Xin-Gao Gong,Shiyou Chen
摘要
ABSTRACT The formation energy of charged defects is a key quantity for assessing dopability and electronic properties of various semiconductor materials, but its evaluation using density functional theory (DFT) is computationally demanding for large‐scale systems. Although machine‐learning models have been introduced to accelerate charged‐defect calculations, limited size transferability when trained only on small defect supercells leads to substantial training costs, and charged‐defect corrections remain a nontrivial challenge. Here, we introduce a machine‐learning Hamiltonian (MLH) model–based method that extends the model's size transferability of electronic structures to charged‐defect formation energies and enables accurate charged‐defect corrections through reliable electrostatic potentials in large supercells. Using oxygen vacancies in SiO 2 and nitrogen vacancies in GaN as benchmarks, a model trained on defect structures with at most 95 atoms accurately predicts formation energies in supercells larger than 400 atoms, with errors below 0.07 eV for neutral defects and below 0.3 eV for +1 and +2 charge states. These results demonstrate the feasibility of extending our approach to more complex defect systems.
科研通智能强力驱动
Strongly Powered by AbleSci AI