Improving Robustness and Training Efficiency of Machine-Learned Potentials by Incorporating Short-Range Empirical...

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Machine learning force fields (MLFFs) are crucial for molecular dynamics simulations.
• Current MLFFs struggle with accuracy and robustness due to limited training data, especially for rare events.
• This deficiency hinders reliable, long-timescale simulations of complex materials.

Purpose of the Study:

• To develop a more robust and training-efficient MLFF framework.
• To address the limitations of purely data-driven MLFFs in capturing essential short-range interactions.
• To enable stable, long-timescale simulations for materials like solid electrolytes.

Main Methods:

• Implemented a hybrid MLFF by integrating an empirical short-range repulsive potential.
• Utilized lithium lanthanum zirconium oxide (Li$_{7}$La$_{3}$Zr$_{2}$O$_{12}$ or LLZO) as a model system.
• Compared the performance of the hybrid MLFF against purely data-driven MLFFs in extended simulations.

Main Results:

• Purely data-driven MLFFs exhibited unphysical atom clustering in LLZO simulations due to insufficient short-range repulsion.
• The hybrid MLFF successfully prevented these artifacts, enabling stable, long-time molecular dynamics simulations.
• The hybrid approach demonstrated high performance with minimal training data (25 configurations) and reduced active learning needs.

Conclusions:

• The hybrid MLFF framework offers a universal paradigm for developing robust and efficient force fields for complex materials.
• Integrating physics-driven constraints with data-driven flexibility enhances MLFF reliability.
• This approach is compatible with existing MLFF architectures and critical for accurate materials modeling.