Scalable Graphene Defect Prediction Using Transferable Learning.

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Graphene possesses exceptional thermal and mechanical properties, making it valuable for advanced technologies like flexible electronics and supercapacitors.
• Material defects significantly degrade graphene's performance, necessitating accurate defect prediction methods.
• Existing machine learning models often require training and prediction data to match in size, limiting their real-world applicability.

Purpose of the Study:

• To develop a transferable machine learning approach for predicting defects in graphene structures of varying sizes.
• To overcome the limitations of previous machine learning techniques that require identical data and material system sizes.

Main Methods:

• A transferable learning approach using logistic regression was developed.
• The model was trained on local vibrational energy distributions from molecular dynamics simulations of small graphene samples.
• The hypothesis was that vibrational energy distributions correlate with local structural anomalies.

Main Results:

• The machine learning model achieved up to 80% accuracy in predicting defects in larger graphene samples.
• The model demonstrated effectiveness on graphene with sizes and shapes not encountered during training.
• The approach proved successful under various practical evaluation metrics.

Conclusions:

• The developed transferable learning approach enables scalable and accurate graphene defect prediction.
• This method facilitates data-driven defect detection for diverse two-dimensional materials.
• The findings open new avenues for utilizing machine learning in materials science for defect analysis.