Scalable machine learning approach to light induced order disorder phase transitions with ab initio accuracy

Area of Science:

• Computational Materials Science
• Condensed Matter Physics
• Photochemistry

Background:

• Machine learning (ML) accurately simulates material thermal properties but struggles with non-thermal phase transitions.
• Accurately describing potential energy surfaces, forces, and vibrational properties under photoexcitation is challenging.
• Simulating light-induced order-disorder transitions requires capturing effects of photoexcited electron-hole plasma.

Purpose of the Study:

• To develop a novel computational approach for simulating light-induced non-thermal phase transitions.
• To create reliable interatomic potentials that account for electron-hole plasma effects on structural properties.
• To investigate the mechanism of non-thermal melting in photoexcited silicon.

Main Methods:

• Combined constrained density functional theory (DFT) with machine learning (ML) to generate interatomic potentials.
• Developed ML potentials capable of capturing electron-hole plasma effects on material properties.
• Performed molecular dynamics (MD) simulations on photoexcited silicon using the developed ML potentials.

Main Results:

• The ML potentials accurately reproduced the phonon dispersion of crystal silicon.
• Simulations involved tens of thousands of atoms, enabling large-scale analysis.
• Identified a soft phonon mode and double-well potential driving non-thermal melting at low temperatures, differing from first-order thermal melting.

Conclusions:

• The novel DFT-ML approach provides highly reliable interatomic potentials for simulating photoexcited materials.
• The findings offer a new understanding of light-induced order-disorder phase transitions, distinct from thermal melting.
• This method enables large-scale, long-time simulations of light-induced phase transitions with ab initio accuracy.