Environment model construction toward auto-tuning of quantum dot devices based on model-based reinforcement learning

Semiconductor quantum dots (QDs) are promising hosts for quantum computers because of their scalability. In order to expedite the development process, there is a strong need for fully automated tuning of QDs that allows for en masse characterization of newly fabricated devices and control over large-scale systems with appreciable variability. Machine learning has been actively explored as a means to this end; however, challenges remain in terms of versatility for different tasks and device types. In this study, we explore a model-based reinforcement learning (MBRL) approach: unlike traditional reinforcement learning techniques, the learning process of MBRL progresses by constructing a model for the environment, which is to be diverted for other tasks and/or devices, thereby minimizing time-consuming learning processes. Using pre-measured data, we construct an environment model and, despite the intrinsic sparse reward distribution of the QD system, demonstrate its suitability for MBRL by emulating the process of auto-tuning to a single QD region. Our results highlight the potential of MBRL for more generic QD auto-tuning techniques, providing a promising step toward fully automated QD tuning.