Combining phonon accuracy with high transferability in Gaussian approximation potential models.

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Atomistic Simulations

Background:

• Machine learning interatomic potentials, like Gaussian Approximation Potential (GAP) models, are increasingly vital for atomistic simulations.
• Accurately predicting vibrational properties across diverse material configurations remains a challenge.

Purpose of the Study:

• To develop a method for fitting GAP models that accurately predict vibrational properties in specific configuration spaces.
• To ensure flexibility and transferability of these models to other configurations.

Main Methods:

• Implemented an adaptive regularization technique for GAP fitting, scaling with atomic force magnitude.
• Interpreted regularization within a Bayesian framework as 'expected error'.
• Tested the approach on structurally diverse silicon allotropes and demonstrated transferability to liquid and amorphous silicon.

Main Results:

• Achieved excellent predictions of phonon modes (0.1 THz–0.2 THz) for various silicon structures.
• Demonstrated high transferability across different material configurations by coupling with existing fitting databases.
• Validated the effectiveness of adaptive regularization for improving GAP model accuracy.

Conclusions:

• The adaptive regularization method provides accurate vibrational property predictions for GAP models.
• This approach enhances the flexibility and transferability of machine learning potentials in materials modeling.
• The developed workflows are beneficial for general GAP-driven materials simulations.