Cluster-Graph Fingerprinting: A Framework for Quantitative Analysis of Machine-Learned Interatomic Model Training and...

Area of Science:

• Computational Materials Science
• Machine Learning in Chemistry and Physics

Background:

• Machine-learned interatomic models (ML-IAMs) significantly advance simulation capabilities, enabling predictions at unprecedented scales.
• The data-driven nature of ML-IAMs introduces challenges in detecting subtle errors, potentially compromising prediction accuracy.

Purpose of the Study:

• To introduce a novel fingerprinting approach for robust analysis of ML-IAM training and application data.
• To enable efficient and statistically rigorous assessment of system configurations within ML-IAM workflows.

Main Methods:

• Development of a graph-based descriptor utilizing the Chebyshev Interaction Model for Efficient Simulation (ChIMES).
• Implementation of a fingerprinting strategy for analyzing system configurations in ML-IAMs.
• Application of the method to assess configuration novelty and dissimilarity.

Main Results:

• The ChIMES-based fingerprints provide efficient and statistically sound analysis of ML-IAM data.
• Demonstrated effectiveness in assessing the novelty of configurations against existing datasets.
• Successfully determined dissimilarity among individual configurations for improved data management.

Conclusions:

• The novel fingerprinting approach addresses critical challenges in ML-IAM accuracy and reliability.
• This method is crucial for enhancing active learning, data curation, and uncertainty quantification in ML-IAM applications.
• Facilitates more trustworthy and efficient use of advanced simulation techniques in scientific research.