Revisiting Machine Learning Potentials for Silicate Glasses: The Missing Role of Dispersion Interactions

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Machine learning interatomic potentials (MLIPs) are emerging as powerful tools for materials simulation.
• Traditional methods face limitations in simulating complex systems like oxide glasses.
• Pretrained MLIPs offer potential for broader applicability, but their performance on disordered systems needs evaluation.

Purpose of the Study:

• To evaluate the pretrained MACE (Multi-ACE) model for simulating silicate glasses.
• To compare MACE performance against a custom-trained DeePMD MLIP for sodium silicates.
• To investigate the impact of dispersion interactions on MLIP accuracy in glass simulations.

Main Methods:

• Utilized the pretrained MACE model and a custom-trained DeePMD MLIP.
• Simulated sodium silicate glasses as a test case.
• Incorporated D3(BJ) dispersion corrections to assess their influence.

Main Results:

• MACE accurately reproduced structural properties like neutron structure factors, pair distribution functions, and Si speciation.
• MACE showed slightly lower accuracy for elastic properties compared to the custom MLIP.
• Both MLIPs demonstrated improved density and elastic property reproduction with dispersion corrections.

Conclusions:

• The pretrained MACE model is suitable for simulating sodium silicate glasses.
• Dispersion interactions are crucial for accurate modeling of glass density and elastic properties.
• Transferability of general MLIPs to disordered systems is feasible but requires careful consideration of training data, especially regarding dispersion.