Learning to shine: Neuroevolution for optical control of phase transitions

The integration of machine learning with experimental and computational approaches is opening new frontiers in controlling complex physical phenomena, such as light-induced phase transitions in quantum materials. Ultrafast structural changes via coherent phonon excitation using intense laser pulses are being explored to induce nonthermal states and nonequilibrium phase transition pathways. However, current methods for discovering stable illumination protocols largely rely on serendipity or trial-and-error approaches. The lack of a systematic framework to consistently design protocols for achieving desired phases hinders progress in the inverse design of non-equilibrium states using light. We present a reinforcement learning framework aimed at optimizing the time-dependent amplitude of the electric field to control optically driven phase transitions. This approach is applied to the response of a prototypical broken symmetry material, bismuth (Bi), under continuous wave and pulsed laser excitations. The system's behavior is modeled using an effective Langevin equation of motion. In both the cases, our approach extends control well beyond the harmonic regime, demonstrating that Fourier Neural Networks can learn and generate optimized illumination protocols for phonon amplification near phase transitions. Additionally, we discuss the experimental applications of this method, showing that the ratio of half-periods in transient reflectivity or absorption oscillations can serve as an effective order parameter for phase transitions. By combining machine learning with experimental methods, this approach enables precise control of phase transitions, offering a robust pathway for the inverse design of non-equilibrium states in quantum materials.