Defects in amorphous materials are the main sources of reliability degradation in modern semiconductor devices. Although density functional theory (DFT) can accurately characterize these defects, its computational cost is extremely high. This work presents a data-efficient approach based on machine learning interatomic potential (MLIP) that significantly reduces computational cost while maintaining high accuracy. We use oxygen vacancy (${\mathrm{V}}_{\text{O}}$) defects on 144 oxygen (O) sites in a 216-atom amorphous ${\mathrm{SiO}}_2$ (a-${\mathrm{SiO}}_2$) supercell as a case study. Twenty different ${\mathrm{V}}_{\text{O}}$ configurations are generated on five O sites and used for training MLIP. The resulting MLIP enables fast structural relaxation of all 576 ($144\times 4$) ${\mathrm{V}}_{\text{O}}$ configurations, achieving a mean absolute error of 0.898 meV/atom, with 540 configurations (93.8%) showing energy errors below 0.3 eV relative to DFT benchmarks. To further improve the accuracy, a dual-model cross-validation strategy is introduced to identify the configurations that MLIP cannot describe accurately. By performing DFT calculations for these flagged configurations, the overall accuracy is improved to 96.5%. This approach reduces the number of heavy supercell DFT calculations from 576 to about 44, effectively balancing accuracy and efficiency and making it ideal for large-scale statistical studies of defects in amorphous materials.

Close

PrevNext

View figure in article

• Defects
• Amorphous materials
• First-principles calculations
• Machine learning