Transferable Neural Network Potentials and Condensed Phase Properties.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Transferable neural network potentials (NNPs) are rapidly advancing, primarily trained on single-molecule data.
• Current applications often focus on molecular simulations, not complex condensed phases.
• The accuracy of NNPs for systems outside their training data is largely unexplored.

Purpose of the Study:

• To evaluate the performance of transferable NNPs in reproducing condensed phase properties.
• To identify specific weaknesses of NNPs when applied to liquid simulations.
• To inform the selection of appropriate NNPs for complex simulations.

Main Methods:

• Assessed two transferable NNPs (ANI-2x, MACE-OFF23(S/M)) against reference data.
• Simulated properties of pure liquids: density, heat of vaporization, heat capacity, isothermal compressibility.
• Analyzed radial distribution functions and self-diffusion constants.

Main Results:

• Both NNPs exhibited specific weaknesses when simulating condensed phases.
• Even minor model flaws led to significant performance degradation in liquid simulations.
• Performance varied considerably between the tested NNP models.

Conclusions:

• Current transferable NNPs may not be sufficiently accurate for general condensed phase simulations.
• Careful model selection and rigorous testing are crucial for NNP applications beyond their training domain.
• Further development is needed to improve NNP reliability for complex systems.