Emergence of accurate atomic energies from machine-learned noble-gas potentials

Area of Science:

• Quantum Chemistry
• Computational Materials Science
• Machine Learning

Background:

• The quantum theory of atoms in molecules (QTAIM) provides local atomic energies, valuable for atomistic machine learning.
• Accurate calculation of these local energies for large systems is computationally expensive.

Purpose of the Study:

• To develop a computationally moderate method for obtaining well-defined local atomic energies.
• To investigate the ability of machine-learned models to reproduce these local energies.
• To assess the transferability of machine-learned models trained on energy decomposition.

Main Methods:

• Utilized semiempirical correlations between total energy components to approximate local atomic energies.
• Applied the method to noble liquids (argon, krypton, and their mixture).
• Trained machine-learned models (neural networks) on total energies and evaluated their ability to predict local energies.

Main Results:

• Achieved well-defined local energies at a moderate computational cost.
• Demonstrated that smaller neural networks trained solely on total energies better reproduce local energy partitioning.
• Showed that models adept at energy decomposition exhibit improved transferability to new systems.

Conclusions:

• A computationally feasible approach to local energy calculation is presented.
• Neural network architecture and training data significantly impact the faithful reproduction of physical properties like local energies.
• Understanding physics within machine learning models enhances their predictive power and transferability.