Accelerating Metadynamics-Based Free-Energy Calculations with Adaptive Machine Learning Potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Increasing demand for accurate free-energy calculations in materials science.
• Ab initio molecular dynamics (AIMD) is computationally expensive.
• Machine learning potentials (MLPs) offer an alternative but require extensive training.

Purpose of the Study:

• To develop an efficient method for free-energy calculations using MLPs.
• To accelerate metadynamics (MetaD) simulations with MLPs.
• To enable easy and fast free-energy surface reconstruction.

Main Methods:

• Proposed adaptive machine learning potential-accelerated metadynamics (AMLP-MetaD).
• MLP (Gaussian approximation potential - GAP) adapts based on uncertainty estimation.
• Integration with reference methods (e.g., density functional theory) for retraining.

Main Results:

• Achieved a 10-time speedup compared to traditional AIMD.
• Obtained free-energy landscapes comparable to ab initio calculations.
• Improved accuracy using Δ-MLP (GAP-corrected density functional tight binding).

Conclusions:

• AMLP-MetaD significantly accelerates free-energy calculations.
• The method is applicable to both cluster and periodic systems.
• Demonstrated successful application in CO adsorption on Pt systems.