Maximizing efficiency of dataset compression for machine learning potentials with information theory.

Area of Science:

• Computational materials science
• Machine learning in chemistry and physics

Background:

• Machine learning interatomic potentials (MLIPs) offer a balance between accuracy and computational cost for materials simulations.
• MLIP performance is highly dependent on the size and diversity of training datasets, with challenges in managing large datasets or ensuring adequate representation in smaller ones.

Purpose of the Study:

• To develop an information-theoretical framework and algorithm for efficient atomistic dataset compression.
• To identify the smallest subset of structures that maximizes information retention and prunes redundancy.

Main Methods:

• Framing dataset compression as a minimum set cover (MSC) problem over atom-centered environments.
• Developing an algorithm to efficiently solve the MSC problem for atomistic datasets.
• Validating the MSC approach on multiple datasets, including GAP-20, TM23, and the ColabFit repository.

Main Results:

• The MSC method effectively retains outliers and preserves dataset diversity, even at high compression rates.
• MLIPs trained on MSC-compressed datasets show reduced error for out-of-distribution data, especially in low-data regimes.
• The MSC approach outperforms conventional subsampling and dimensionality reduction methods in preserving dataset information and improving MLIPs.

Conclusions:

• The developed MSC algorithm provides an efficient and effective method for atomistic dataset compression.
• This approach enhances the accuracy and reliability of MLIPs by optimizing training data selection.
• The QUESTS package implementation facilitates broader application in atomistic modeling, including data subsampling and outlier detection.