Machine Learning Potentials with the Iterative Boltzmann Inversion: Training to Experiment.

Area of Science:

• Computational materials science
• Machine learning in physics
• Quantum mechanical simulations

Background:

• Machine learning potentials (MLPs) trained on quantum-mechanical data show great promise.
• Integrating experimental data into existing MLP training is challenging due to data heterogeneity.
• Current methods struggle to combine simulated and experimental data effectively.

Purpose of the Study:

• To develop a training procedure for MLPs that incorporates experimental data.
• To improve the accuracy of MLPs for molecular dynamics simulations.
• To address limitations in current MLP training methodologies.

Main Methods:

• Investigated a training procedure based on iterative Boltzmann inversion.
• Generated pair potential corrections to existing MLPs using equilibrium radial distribution function data.
• Applied corrections to a density functional theory-based MLP for pure aluminum.

Main Results:

• The corrected MLP significantly reduced overstructuring in the melt phase of aluminum.
• The enhanced MLP demonstrated improved prediction of experimental diffusion constants.
• The method avoids complex procedures like autodifferentiating through molecular dynamics solvers.

Conclusions:

• A practical framework for integrating experimental data into MLPs was presented.
• The developed method enhances the accuracy of molecular dynamics simulations.
• This approach offers a viable way to leverage diverse data sources for materials modeling.