An Active Regression Learning Method for Quality Evaluation of Molecular Dynamics Data

Training deep potentials for molecular dynamics simulations via active learning methods is a prevalent technique. However, a significant drawback of this method is its tendency to get trapped in confined regions during conformational searches, potentially resulting in datasets with redundant information. In this research, we propose the integration of active regression learning to construct regression models that link data energy and its distribution to quality scores. This approach enables us to assess the quality of each data entry within the dataset. We have validated our method on three datasets, and the outcomes demonstrate its exceptional proficiency in eliminating lower-quality data points from molecular simulation datasets.