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

We present a new method for fingerprint- ing atomic configurations relevant to ML-IAM training and application, utilizing the ChIMES descriptor. These fingerprints enable rigor- ous analysis of statistical distinguishability be- tween configurations. Sample applications in- clude assessing diversity within ML-IAP train- ing datasets, monitoring structural equilibra- tion during simulations, and automating the monitoring of ML-IAM active learning work- flows. Ultimately, these fingerprints can be de- ployed in tasks aimed at enhancing ML-IAM robustness and reliability, such as automated training dataset curation, active learning, and uncertainty quantification.