Uncertainty estimation for molecular dynamics and sampling

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Statistical Mechanics

Background:

• Machine-learning potentials (MLPs) accelerate simulations by replacing expensive electronic-structure calculations.
• MLP accuracy relies on training data; predictions outside the training space are unreliable.
• Uncertainty in individual configurations propagates to thermodynamic averages, limiting simulation accuracy in unexplored regions.

Purpose of the Study:

• To develop methods for quantifying uncertainty in machine-learning potential predictions.
• To enhance the resilience and accuracy of molecular dynamics simulations using MLPs.
• To support active learning strategies for efficient model development.

Main Methods:

• Utilized uncertainty quantification (UQ) with baseline energy models or less accurate interatomic potentials.
• Introduced an on-the-fly reweighing scheme for estimating uncertainty in thermodynamic averages from long trajectories.
• Applied methods to diverse systems (water, liquid gallium) and properties (structural, thermodynamic).

Main Results:

• Demonstrated that UQ improves the robustness of simulations, especially in unexplored regions of phase space.
• The on-the-fly reweighing scheme effectively estimates uncertainty in thermodynamic averages.
• Validated the approach across different material systems and property types.

Conclusions:

• Uncertainty quantification is crucial for reliable machine-learning potential-based simulations.
• The proposed methods enhance simulation accuracy and support active learning for model improvement.
• UQ provides a pathway to more trustworthy and efficient computational materials discovery.