Active learning of molecular data for task-specific objectives.

Area of Science:

• Computational Chemistry
• Machine Learning
• Data Science

Background:

• Active learning (AL) is a promising machine learning approach for data efficiency.
• Its practical utility and potential for computational savings are application-dependent.
• Understanding when AL yields benefits is crucial for practitioners.

Purpose of the Study:

• To systematically assess active learning performance across diverse molecular datasets and scientific tasks.
• To identify factors influencing AL's effectiveness and computational savings.
• To provide guidance on optimal AL strategies for molecular data analysis.

Main Methods:

• Implemented active learning using Gaussian processes (GP) with a many-body tensor molecular representation.
• Evaluated AL on tasks of dataset compilation and targeted molecular searches.
• Tested various acquisition strategies, batch sizes, and GP noise settings.

Main Results:

• AL performance varied by task; it excelled in targeted molecular searches, yielding up to 64% data savings.
• For dataset compilation, AL performance was sensitive to GP noise settings and acquisition strategies.
• Optimal AL performance for targeted searches occurred when target molecules had minimal overlap with the overall dataset distribution.

Conclusions:

• Active learning's effectiveness is task-specific, particularly influenced by target and dataset distributions.
• AL offers significant computational savings for targeted molecular searches.
• Careful selection of AL strategies and understanding data distribution are key to maximizing benefits.