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

In materials science, the choice of structural descriptors for machine learning protocols strongly influences both predictive performance and model interpretability. High-dimensional descriptors can improve numerical accuracy, but often introduce substantial computational overhead and reduce transparency. To address this, the Dynamic Collision Fingerprint (DCF) framework generates concise descriptors via the dynamical probing of atomic structures. In this work, we benchmark DCF against the widely used Matminer library using a data set of 120 two-dimensional (2D) carbon allotropes. We evaluate performance across three regression algorithms, linear regression, decision trees, and XGBoost, utilizing train-test partitions from 10% to 90%. Our results demonstrate that DCF matches the predictive accuracy of Matminer across all algorithms. Although Matminer can be faster in terms of execution time, DCF accomplishes its predictive performance using descriptors that are significantly lower-dimensional, pointing to manageable computing costs in feature space. Moreover, compared to the rather technical Matminer descriptions, DCF exhibits considerably clearer physical interpretability. These findings suggest that DCF serves as a viable alternative to high-dimensional descriptor libraries for structural representation, since it remains both computationally flexible and physically grounded.