Mitigating semiconductor device variability with machine learning

A concerning consequence of quantum device variability is that the tuning of each qubit in a quantum circuit constitutes a time-consuming non-trivial process that has to be independently performed for each device, requiring a deep understanding of the particular device to be tuned and "muscle memory". I will show machine-learning-based approaches that can tune and characterise quantum devices completely automatically, regardless of the device architecture and the material realisation. Our algorithms are able to tune double quantum dot devices defined in Si FinFETs, Ge/Sicore/shell nanowires, and both SiGe and AlGaAs/GaAs heterostructures, successfully accommodating the different modes of gate operation (depletion/accumulation), disorder and noise characteristics. We report tuning times as fast as 10 minutes starting from scratch – well over an order of magnitude faster than what would be achievable by a dedicated expert human operator. Just as AlphaZero showed that the achievements of AlphaGo could be extended to learning to win at different board games without needing to be reprogrammed for each, so our result shows that control of complex quantum device circuits can be achieved using machine learning.