Accurate thermophysical properties of water using machine-learned potentials.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Simulating water from first principles is computationally intensive due to slow system dynamics.
• Machine-learned interatomic potentials (MLPs) accelerate simulations but often lack accuracy for reliable uncertainty quantification.

Purpose of the Study:

• To evaluate the performance of MACE, an equivariant graph neural network architecture, for simulating water's thermophysical properties.
• To compare the accuracy and efficiency of MACE against traditional methods like kernel-based potentials (KbPs).

Main Methods:

• Training MACE using an extensive RPBE-D3 database.
• Predicting density isobars, diffusion constants, radial distribution functions, and melting points using MACE.
• Validating MACE predictions against ground-truth density functional theory (DFT) ensembles.

Main Results:

• Equivariant MACE models achieved significantly lower total energy errors compared to KbPs.
• MACE enabled reliable thermodynamic reweighting with minimal bias.
• Validation against DFT ensembles was achieved with increased efficiency using MACE.

Conclusions:

• Equivariant MLPs, specifically MACE, provide DFT-level accuracy for simulating water's thermophysical properties.
• MACE represents a robust and reliable advancement for computational materials science.
• These findings highlight the potential of equivariant MLPs for accelerating complex molecular simulations.