Neural network representation of electronic structure from ab initio molecular dynamics

Area of Science:

• Computational Materials Science
• Condensed Matter Physics
• Machine Learning in Physics

Background:

• Ab initio molecular dynamics simulations generate vast electronic structure data, often underutilized.
• Accurate electronic structure is crucial for understanding material properties and dynamics.

Purpose of the Study:

• To introduce a transferable, high-fidelity neural network representation of electronic structure data.
• To enable efficient and accurate electronic evolution and sampling in materials simulations.
• To explore novel physical phenomena in materials like carbyne.

Main Methods:

• Developed a neural network model generating tight-binding Hamiltonians for crystalline materials.
• Integrated the neural network with machine learning-boosted molecular dynamics.
• Applied the model to a 1D charge-density wave material (carbyne) to compute spectral functions and optical conductivity.

Main Results:

• Demonstrated efficient and accurate electronic structure representation and dynamics.
• Computed spectral functions and optical conductivity for carbyne.
• Revealed significant renormalization of low-energy modes during soliton-antisoliton pair annihilation due to electron-lattice coupling beyond the Born-Oppenheimer limit.

Conclusions:

• The developed surrogate model provides an efficient and reusable tool for electronic structure dynamics.
• Enables the calculation of complex physical properties, opening new avenues in materials modeling.
• Facilitates the study of phenomena beyond the Born-Oppenheimer approximation.