Transferable Water Potentials Using Equivariant Neural Networks.

Area of Science:

• Computational chemistry
• Materials science
• Statistical mechanics

Background:

• Machine learning interatomic potentials (MLIPs) offer a cost-effective alternative to quantum mechanics for molecular simulations.
• Previous studies suggested MLIPs trained solely on liquid water data lack transferability to other phases, like vapor-liquid equilibrium.
• This limitation implies MLIPs may not capture the fundamental physics of water interactions.

Purpose of the Study:

• To develop and validate MLIPs with equivariant architecture for accurate water simulations across different phases.
• To assess the transferability of these MLIPs to vapor-liquid equilibrium, gas-phase clusters, and solid ice phases.
• To confirm if equivariant MLIPs can learn physically accurate water interactions.

Main Methods:

• Training machine learning interatomic potentials (MLIPs) using an equivariant architecture on 3200 liquid water structures.
• Validating the MLIPs against experimental and theoretical data for liquid water properties (density).
• Testing the MLIPs' performance in simulating vapor-liquid equilibrium, gas-phase water clusters (many-body decomposition), and ice phases (energetics, vibrational density of states).

Main Results:

• The developed equivariant MLIPs accurately reproduced liquid water density across a wide temperature range (230-365 K).
• The MLIPs demonstrated excellent transferability, accurately predicting vapor-liquid equilibrium properties up to 550 K.
• The potentials successfully captured many-body interactions in gas-phase water clusters and accurately predicted properties of ice phases.

Conclusions:

• Equivariant machine learning interatomic potentials trained on liquid water exhibit remarkable transferability across arbitrary water phases.
• These MLIPs provide accurate simulations of water behavior, including phase transitions and intermolecular interactions.
• The findings suggest that equivariant architectures enable MLIPs to learn physically meaningful representations of water interactions.