Automatic tuning of quantum dot metrology experiments with noise-aware machine learning

The successful operation of intricate quantum experiments typically relies on a careful fine-tuning of control parameters informed by repeated measurements of an operated quantum device. These time-consuming measurements can often become a significant bottleneck in the screening and iterative improvement of fabricated devices. In this talk, I will discuss a recently introduced approach tackling this task for a fully automated tuning of single-electron pumps, nanoscale quantum dot experiments allowing for the generation of well-quantized currents. It harnesses the power of machine learning to iteratively identify informative points of the control parameter space for an accurate characterization of the device with a small number of measurements. Together with a carefully crafted post-processing approach, this active learning scheme allows us to reliably locate the small operational regimes in the vast parameter landscape and provide accurate quantifications of the single-electron quantization error. We demonstrate our scheme for the characterization of a multiplexed array of GaAs electron pumps, for which we achieve an approximately 8-fold reduction in the time required for the characterization as compared to a conventional semi-automated sweeping through the parameter space. This paves the way for a fully automated high-throughput testing and characterization of multiple devices without requiring additional manual intervention for the large-scale realization of the fundamental definition of the ampere. Building upon these new capabilities, I also discuss new machine learning schemes to "look beyond the noise" of measurement data with Bayesian modelling principles, improving the quality of the device characterization by bringing together noisy measurement data with physical intuition.