Less is more: Sampling chemical space with active learning

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Developing accurate and transferable machine learning (ML) potentials for molecular energetics is challenging.
• Data generation for training ML potentials is not well understood or researched.
• Human bias often influences the selection of new training data, impacting ML potential accuracy.

Purpose of the Study:

• To present a fully automated approach for generating datasets to train universal ML potentials.
• To leverage active learning (AL) via Query by Committee (QBC) for efficient data sampling.
• To mitigate human bias and reduce the training set size for ML potentials.

Main Methods:

• Implemented an active learning (AL) algorithm based on Query by Committee (QBC).
• QBC uses disagreement within an ensemble of ML potentials to identify unreliable predictions.
• Developed the COmprehensive Machine-learning Potential (COMP6) benchmark for validation.

Main Results:

• AL-based ANI potentials achieved superior performance compared to random sampling with significantly less data (10% of data).
• The final AL-based model outperformed the original ANI-1 potential on the COMP6 benchmark after training on only 25% of the data.
• The developed universal ANI-1x potential demonstrated accuracy comparable to state-of-the-art ML potentials across the COMP6 benchmark.

Conclusions:

• The automated AL approach effectively generates high-quality datasets for training universal ML potentials.
• This method significantly reduces the data needed for training, outperforming traditional random sampling.
• The resulting universal ML potential (ANI-1x) offers accurate energy and force predictions for organic molecules containing CHNO elements.