Digitally stimulated Raman passage by deep reinforcement learning

Preparing an arbitrary preselected coherent superposition of quantum states finds widespread application in physics, including initialization of trapped ion and superconductor qubits in quantum computers. Both fractional and integer stimulated Raman adiabatic passage involve smooth Gaussian pulses, designed to grant adiabaticity, so to keep the system in an eigenstate constituted only of the initial and final states. We explore an alternative method for discovering appropriate pulse sequences based on deep reinforcement learning algorithms and by imposing that the control laser can be only either on or off instead of being continuously amplitude-modulated. Despite the adiabatic condition is violated, we obtain fast and flexible solutions for both integer and fractional population transfer. Such method, consisting of a Digital Stimulated Raman Passage (D-STIRaP), proves to be particularly effective when the system is affected by dephasing therefore providing an alternative path towards control of noisy quantum states, like trapped ions and superconductor qubits.