Energy-conserving molecular dynamics is not energy conserving.

Area of Science:

• Computational Chemistry and Physics
• Materials Science
• Machine Learning Applications

Background:

• Molecular dynamics (MD) simulations are crucial for understanding molecular and materials properties.
• Traditional MD simulations prioritize energy conservation, but machine learning (ML) potentials introduce challenges.
• Current ML methods for MD may yield unphysically dissociative dynamics or lack guaranteed energy conservation.

Purpose of the Study:

• To propose a clear distinction between simulation-energy and true-energy conservation in MD.
• To introduce novel, simple criteria for evaluating MD simulation quality based on true-energy non-conservation.
• To establish new standards for assessing MD methods, focusing on physical realism over mere energy conservation.

Main Methods:

• Developed new criteria to estimate the degree of true-energy non-conservation in MD simulations.
• Applied these criteria to assess the quality of MD simulations, particularly those using ML potentials.
• Demonstrated the practical utility of the new criteria using infrared spectra simulations.

Main Results:

• The proposed criteria effectively estimate true-energy non-conservation in MD simulations.
• These criteria offer a more intuitive and important evaluation metric than assessing ML potential energies and forces alone.
• The method is universally applicable, even for trajectories with unknown or discontinuous potential energy landscapes.

Conclusions:

• Focusing on true-energy conservation provides a more accurate measure of MD simulation quality.
• The new criteria offer a practical and universal approach to evaluating MD simulations.
• This work can guide the development of more accurate computational methods for molecular and materials simulations.