Deep Learning for Molecular Control in Noisy, Partially Observable Environments

Quantum-logic spectroscopy (QLS) has become an exciting field of discovery, with potential applications to precision measurements for detection of symmetry violations. Especially for high resolution measurements of molecular spectra and precise control of molecular population, preparing trapped molecular ions in a known pure state is an important problem. It has been demonstrated that the goal can be achieved with repetitive projective measurements with QLS for simple molecules. However, due to experimental imperfections and shifts rising from unintended environmental electric fields, a more effective scheme for fast and robust preparation in noisy experimental environments is needed for more complex molecules. Here we utilize reinforcement learning (RL), an established method for quantum control to address the problem, illustrated by the example of a CaH+ molecular ion co-trapped with a Ca+ ion. We train the RL agent in noisy environments for robustness and compare its performance in environments with unknown dynamics against the conventional scheme utilizing optical pumping. We further survey through various combinations of transformers and RL agents to understand how AI can assist in tackling the problem of fast and high-accuracy pure state preparation with only partial observations of the quantum state.