Random Sampling Versus Active Learning Algorithms for Machine Learning Potentials of Quantum Liquid Water

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate machine learning potentials require comprehensive electronic structure data.
• Data generation is computationally intensive, necessitating efficient structure selection schemes.
• High-dimensional neural network potentials (HDNNPs) are increasingly used for molecular simulations.

Purpose of the Study:

• Compare HDNNPs trained on random sampling versus active learning datasets for quantum liquid water.
• Investigate the impact of data selection strategies on potential accuracy and structural properties.
• Identify optimal methods for constructing training datasets for machine learning potentials.

Main Methods:

• Trained HDNNPs on quantum liquid water using datasets from random sampling and active learning (query by committee).
• Analyzed test errors and structural properties based on different training data generation methods.
• Evaluated the influence of energy offsets and correlations on model performance.

Main Results:

• Random sampling resulted in smaller test errors compared to active learning for a given dataset size.
• HDNNPs trained on as few as 200 structures accurately predicted structural properties of quantum liquid water.
• Energy correlations proved more robust than energy offsets as an error measure.

Conclusions:

• The choice of training data construction algorithm has a limited impact on the final accuracy of HDNNPs for structural properties.
• Active learning requires careful consideration of initial datasets to avoid exploring irrelevant configurations.
• Random sampling offers a competitive alternative to active learning for training machine learning potentials.