Enhanced sampling of robust molecular datasets with uncertainty-based collective variables

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Generating representative datasets is vital for robust machine-learned interatomic potentials (MLIPs).
• Complex molecular systems present challenges due to intricate potential energy surfaces and numerous local minima.
• Traditional data generation methods are often inefficient or fail to capture critical configurations.

Purpose of the Study:

• To develop a data-efficient active learning strategy for generating high-quality datasets for MLIPs.
• To address the challenge of exploring complex potential energy surfaces and overcoming energy barriers.
• To enhance the robustness and accuracy of machine-learned interatomic potentials.

Main Methods:

• Leveraging model uncertainty as a collective variable (CV) to guide data acquisition.
• Employing a Gaussian Mixture Model-based uncertainty metric from a single model.
• Utilizing biased molecular dynamics simulations for targeted exploration of configuration space.

Main Results:

• Demonstrated effectiveness in overcoming energy barriers and exploring previously unseen energy minima.
• Successfully enhanced datasets in an active learning framework.
• Validated the approach on alanine dipeptide and bulk silica systems.

Conclusions:

• The proposed uncertainty-guided method significantly improves data generation efficiency for MLIPs.
• This approach enables more robust and accurate molecular modeling by focusing on chemically relevant, uncertain regions.
• Active learning frameworks can be effectively enhanced by uncertainty quantification for targeted data acquisition.