Condensed Matter Seminar Series

Natalia Ares

University of Oxford

Quantum labs no longer operated by human experts

Machine learning has been the enabler of well-known breakthroughs in computer science, such as the victory of AI over Go world champions, and superhuman face recognition. This maturing technology can be directed to the real-time control of quantum devices. 

Our algorithms harness bespoke machine learning techniques for the fully-automated tuning and characterization of semiconductor devices [1]. Gaussian processes have allowed us to create an algorithm able to tune a ‘virgin’ quantum dot device to operating conditions faster than human experts, without human input and robust across different device architectures and material systems. In the same way a Go player carefully balances short and long-term goals and devises actions accordingly, one of our deep reinforcement learning algorithms devises efficient policies in real time to find desired measurement features. This algorithm can reduce by an order of magnitude the long characterization times imposed by device variability. I will also demonstrate fine tuning of a quantum dot device using unsupervised embedding learning. These approaches are widely applicable, opening the way to a completely automatic and efficient route to quantum device measurement and tuning, and thus taking a crucial step towards the scalability of quantum circuits.


[1] N. Ares, Machine learning as an enabler of qubit scalability, Comment Nature Reviews Materials (2021) 


Zoom-password: NielsBohr