Comparing machine learning potentials for water: Kernel-based regression and Behler-Parrinello neural networks.

Area of Science:

• Computational chemistry
• Materials science
• Statistical mechanics

Background:

• Accurate prediction of thermodynamic properties is crucial for understanding water's behavior.
• Machine learning potentials (MLPs) offer a promising avenue for efficient and accurate simulations.

Purpose of the Study:

• To evaluate the performance of different MLPs in predicting water's thermodynamic properties.
• To assess the impact of dataset size and generation method on MLP accuracy.

Main Methods:

• Kernel-based regression and high-dimensional neural networks were employed.
• MLPs were trained on datasets of varying sizes generated using accurate methods and on-the-fly learning.
• Performance was evaluated by predicting diffusion constants, pair-correlation functions, and density isobars.

Main Results:

• Excellent agreement was observed for diffusion constants and pair-correlation functions, especially with larger datasets.
• Predicted density isobars showed acceptable variations, considering inherent errors in approximate density functional theory.
• Minor differences between MLPs did not significantly impact the prediction of key observables.

Conclusions:

• The quality and size of the training dataset are more critical than the specific MLP fitting method.
• Root mean square errors have limitations; comprehensive testing with multiple MLPs is recommended for complex properties.