Out-of-distribution evaluation of active learning pipelines for molecular property prediction

Key Points

• This research aims to evaluate active learning approaches for predicting molecular properties, focusing on out-of-distribution scenarios.
• Utilized evidential deep learning techniques in active learning frameworks.
• Compared active learning against random sampling methods.
• Evaluated solvation free-energy predictions using a consistent labeling budget.
• Active learning outperformed random sampling, achieving lower test RMSE.
• Improved generalization to previously unseen chemical environments.