Chemical Feature Engineering and Defect-Aware Structural Fingerprint Representations for Complex Defects in 2D...

Area of Science:

• Materials Science
• Computational Materials Science
• Condensed Matter Physics

Background:

• Designing effective descriptors for multiple defects in 2D materials is complex due to varied atomic environments.
• Existing methods struggle to differentiate similar defect compositions and lack interpretability or require extensive data.

Purpose of the Study:

• To engineer improved chemical descriptors and structural features for defect representation in 2D materials.
• To enhance model interpretability, generalization, and computational efficiency for defect property prediction.

Main Methods:

• Engineered chemical descriptors and constructed structural features from nearest-neighbor distributions (Classical Force-Field-Inspired Descriptors - CFID).
• Incorporated defect-aware structural features using Hellinger distance to improve data point discrimination.
• Evaluated model performance on predicting formation energy and HOMO-LUMO gap for 2D material defects.

Main Results:

• The engineered features improved data point discrimination by 50% while enhancing model interpretation and generalization for formation energy prediction.
• The model showed superior stability and lower uncertainty on an external dataset of bulk hBN defects.
• Predicting complex targets like the HOMO-LUMO gap remained challenging, with no improvement over baseline descriptors.

Conclusions:

• The physics-informed approach provides an interpretable and computationally efficient alternative to deep learning for defect analysis in 2D materials.
• This method aids in the high-throughput screening of stable defect candidates, reducing the need for costly first-principles calculations.