Locality-Sensitive Hashing-Based Data Set Reduction for Deep Potential Training.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning (ML) potentials require extensive data from expensive ab initio calculations for accuracy.
• Current methods face challenges in balancing data set size, diversity, and computational cost.

Purpose of the Study:

• To develop a novel method for reducing the size of training data sets for ML potentials.
• To decrease the computational expense of ab initio calculations while maintaining data quality and diversity.
• To enable accurate free energy calculations for complex systems.

Main Methods:

• Implemented a locality-sensitive hashing (LSH) approach to select diverse and accurate data points.
• Developed ML potentials using the reduced data sets.
• Performed well-tempered metadynamics simulations with the ML potentials.

Main Results:

• Achieved data set size reductions of nearly an order of magnitude.
• Successfully developed ML potentials for a chemical reaction and a phase transition.
• Calculated converged free energy surfaces for both systems.

Conclusions:

• The LSH-based method effectively minimizes data requirements for ML potentials.
• This approach significantly reduces the cost of developing accurate ML potentials.
• Enables efficient free energy calculations for chemical and material processes.