Using deep learning for digitally controlled STIRAP

Stimulated Raman Adiabatic Passage (STIRAP) is a technique for preparing atoms and molecules in an arbitrary preselected coherent superposition of quantum states, ordinarily controlled by Gaussian laser pulses. We search novel pulse sequences by exploiting deep reinforcement learning algorithms in order to achieve fast and flexible solutions for integer and fractional STIRAP. By using the robustness of the PPO algorithm, we impose the agent to exploit only digital pulses corresponding to on/off states of the control radiation instead of continuous amplitudes. Such method allows to adapt to detuning of the energy levels and disturbances such as dephasing.