Comparative assessment of composition- and structure-based surrogate models across 2D materials databases.

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning Applications

Background:

• Machine learning (ML) surrogate models accelerate materials discovery.
• Their transferability across diverse 2D materials databases is not well understood.

Purpose of the Study:

• Benchmark ML surrogate model accuracy and transferability across three major 2D materials databases.
• Investigate factors influencing model performance during cross-database transfer.

Main Methods:

• Evaluated composition-based and structure-based surrogate models on C2DB, 2DMatPedia, and JARVIS-2D databases.
• Assessed predictive performance for energy per atom and bandgap.
• Analyzed effects of dataset size and coverage via down-sampling and error correlations.

Main Results:

• Energy predictions were robust; bandgap predictions were challenging due to data imbalance and DFT parameter inconsistencies.
• Composition-based models demonstrated more stable cross-database performance than structure-based models.
• Training dataset coverage, feature representation, and model architecture significantly impact transferability.

Conclusions:

• Cross-database performance of ML surrogate models in 2D materials discovery is primarily governed by training data coverage and model specifics.
• Composition-based models provide more reliable generalization across heterogeneous materials datasets.
• Understanding these factors is crucial for robust ML-driven materials discovery.