Hyperparameter Optimization for Atomic Cluster Expansion Potentials.

Area of Science:

• Computational Materials Science
• Materials Informatics
• Machine Learning in Chemistry

Background:

• Machine learning (ML) interatomic potentials are crucial for simulating materials at large scales.
• The Atomic Cluster Expansion (ACE) framework offers promising performance for developing accurate ML potentials.

Purpose of the Study:

• To present a largely automated computational approach for optimizing hyperparameters in ACE potential models.
• To extend the open-source Python package XPOT with an interface for ACE fitting.
• To demonstrate the impact of hyperparameter selection on model performance.

Main Methods:

• Systematic hyperparameter sweeps to optimize the functional form and complexity of ACE models.
• Integration of ACE fitting capabilities into the XPOT Python package.
• Validation of the optimization approach on silicon and Sb2Te3 systems.

Main Results:

• Demonstrated a largely automated workflow for optimizing ACE potential hyperparameters.
• Successfully extended the XPOT package to facilitate ACE model development.
• Showcased the effectiveness of the approach for covalent (Si) and phase-change (Sb2Te3) materials.

Conclusions:

• Hyperparameter selection is critical for developing high-performance ML potential models.
• The automated approach and extended XPOT package streamline the creation of accurate interatomic potentials.
• This work advances the field of ML-driven materials discovery and simulation.