A Comparative Study of Structural Representations for 2D Materials: Insights from Dynamic Collision Fingerprint and...

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning in Materials

Background:

• Structural descriptors are crucial for machine learning in materials science, impacting predictive accuracy and interpretability.
• High-dimensional descriptors offer accuracy but incur computational costs and reduce transparency.

Purpose of the Study:

• To benchmark the Dynamic Collision Fingerprint (DCF) against the Matminer library for representing atomic structures.
• To evaluate descriptor performance across various machine learning regression algorithms and data set sizes.

Main Methods:

• Utilized a dataset of 120 2D carbon allotropes.
• Compared DCF with Matminer using linear regression, decision trees, and XGBoost.
• Evaluated performance across train-test partitions ranging from 10% to 90%.

Main Results:

• DCF achieved predictive accuracy comparable to Matminer across all tested algorithms.
• DCF employs significantly lower-dimensional descriptors than Matminer, indicating computational efficiency.
• DCF descriptors demonstrated superior physical interpretability compared to Matminer.

Conclusions:

• DCF presents a computationally flexible and physically grounded alternative to high-dimensional descriptor libraries.
• The framework offers a viable option for structural representation in machine learning applications.
• DCF balances predictive performance with enhanced model transparency and reduced computational overhead.