Learning to shine: Neuroevolution enables optical control of phase transitions

Abstract

We address active optical steering of structural phase transitions in solids. We demonstrate that existing evolutionary reinforcement learning approaches can derive optimal time-dependent electric fields in driven dissipative classical systems far beyond the harmonic regime, enabling the stabilization of non-thermal structural phases. Our approach relies on experimentally-extractable metrics of the phase-space evolution and interpretable Fourier Neural Network surrogates of the electric field. Using this method on first-principles models, we obtain protocols stabilizing a symmetric phase in bismuth through impulsive Raman and displacive excitations with continuous and pulsed light sources in the presence of dissipation and thermal disorder. The method is gradient-free, enabling optimization loops based solely on experimental data and providing a practical route for controlling light-induced structural dynamics independently of the microscopic model of light-matter interactions.