A machine learning approach to light-induced order-disorder phase transitions: large-scale long-time simulations with ab initio accuracy.

While machine learning has excelled in simulating material thermal properties, its application to order-disorder non-thermal phase transitions induced by visible light has been limited by challenges in accurately describing the potential energy surface, the forces and the vibrational properties in the presence of a photoexcited electron-hole plasma. Here, we present a novel approach that combines constrained density functional perturbation theory with machine learning techniques, yielding highly reliable interatomic potentials capable of capturing electron-hole plasma effects on structural properties. Applied to photoexcited silicon, our potential is ten times more accurate than previous models and allows for simulations up to tens of thousands of atoms. We show that, at low enough temperatures, the non-thermal melting transition is driven by a soft phonon and the formation of a double well potential, at odds with thermal melting being strictly first order.

Our method paves the way to large-scale long-time simulations of light-induced order-disorder phase transitions in materials with {\it ab initio} accuracy.