Limitations of Cluster-Trained MLIPs for Liquid Density and Diffusivity.

Area of Science:

• Computational chemistry and materials science.
• Development and application of machine-learned interatomic potentials (MLIPs).

Background:

• Machine-learned interatomic potentials (MLIPs) bridge classical force-field efficiency with quantum mechanical accuracy.
• Accurate simulation of liquid properties is crucial for battery electrolyte development.

Purpose of the Study:

• To evaluate uncertainties in MACE-architecture MLIPs trained on different datasets (periodic vs. molecular clusters).
• To assess the impact of MLIP uncertainties on simulated liquid densities and diffusivities for battery solvents.
• To determine if MLIP uncertainties are small enough for comparing density functional theory (DFT) functionals.

Main Methods:

• Training MACE-based MLIPs using two distinct public datasets: one with periodic structures, another with molecular clusters.
• Performing molecular dynamics (MD) simulations using trained MLIPs to calculate density and diffusivity.
• Analyzing model uncertainties by varying regression seeds and training set sizes.

Main Results:

• All trained MACE-MLIPs produced stable 1 ns NPT trajectories for battery solvents.
• MLIPs trained on molecular cluster data exhibited significant sensitivity to training seeds and data selection.
• Cluster-trained MLIPs resulted in substantial uncertainties for simulated density and diffusivity.

Conclusions:

• MLIPs trained on periodic structures provide more reliable simulations of liquid properties compared to those trained on cluster data.
• Uncertainties in cluster-trained MLIPs are too large for robust comparison of DFT functionals.
• The developed workflow can be extended for broader assessment of DFT functionals and MLIP performance.