Efficient quantum device tuning using machine learning

Lead Research Organisation: University of Oxford
Department Name: Materials

Abstract

Fault-tolerant quantum computers require hundreds to millions of physical qubits to be operated with high fidelity. Inevitable hardware imperfections must be tuned away through iterative interplay of characterization, simulation, and parameter refinement, with each data point informing the decision of what to measure next. The technology is only scalable if this task can be efficiently automated. In the language of computer science, this is a Bayesian optimization problem. Recent progress in machine learning, currently one of the most rapidly developing fields of computing, makes it possible to automate the entire process. This project will apply these new techniques experimentally, working with leaders in machine learning.

Fault-tolerant quantum computers require hundreds to millions of physical qubits to be operated with high fidelity. Inevitable hardware imperfections must be tuned away through iterative interplay of characterization, simulation, and parameter refinement, with each data point informing the decision of what to measure next. The technology is only scalable if this task can be efficiently automated. In the language of computer science, this is a Bayesian optimization problem. Recent progress in machine learning, currently one of the most rapidly developing fields of computing, makes it possible to automate the entire process. This project will apply these new techniques experimentally, working with leaders in machine learning.

The focus will be electron spin qubits in gate-defined GaAs quantum dots. These are an ideal testbed because the physics is known and the dot potential and tunnel barriers are conveniently optimized by tuning gate voltages. Nonetheless, tuning a simple device by hand takes days to weeks, which is clearly not scalable.

The goal of this project is to develop a machine to automatically tune a singlet-triplet qubit in a double quantum dot. The machine will use electrical measurements of the quantum dot to deduce device parameters in the most efficient way, and then adjust gate voltages to optimise them. Inevitably, device imperfections lead to trade-offs in how it is tuned, and we will use simulations of small qubit clusters to identify how to optimise these to make spin qubits useful even in the presence of errors.

There has been rapid progress in semiconductor spin qubit devices, but in most cases these were tuned by hand (although using a computer to control data sweeps). There has been at least one pioneering experiment, at TU Delft, using a feature recognition algorithm to tune up a double quantum dot. This project aims to be the first to use the full power of a Bayesian approach to estimate and optimise device parameters.

This project makes use of new EPSRC-funded facilities in Oxford, including a facility for hardware-in-the-loop testing of quantum technology, and the NQIT computing facility. The applications of this approach will, we hope, ultimately extend to many areas of experimental science. The project aligns most closely with the EPSRC priority areas of "Quantum devices components and systems" and "Artificial intelligence technologies"

EPSRC's research areas:

Physical sciences
Quantum technologies

Publications

10 25 50
publication icon
Lennon D (2019) Efficiently measuring a quantum device using machine learning in npj Quantum Information

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509711/1 01/10/2016 30/09/2021
1944854 Studentship EP/N509711/1 01/10/2017 30/09/2020 Dominic Lennon
 
Description As a result of work funded by this award we have achieved the following:
We have shown how a machine learning algorithm can efficiently measure semiconductor quantum devices, significantly reducing measurement times. The algorithm takes a scientific approach: it starts with a set of partial measurements, makes predictions of what the data might look like, and decides which measurement(s) would be the most informative to perform next. For this purpose, we combined, for the first time, a deep-generative model and an information-theoretic approach. This work has significantly impacted the quantum information and machine learning communities, and has attracted significant press coverage (http://www.ox.ac.uk/news/2019-09-26-machine-learning-helps-open-new-possibilities-quantum-devices).
More recently in collaboration with Google DeepMind, we reported the first demonstration of automatic tuning of quantum devices faster than human experts without human input (currently under review in Nature Communications). We were able to show how the algorithm can dynamically tune a 'virgin' semiconductor device to operation conditions. This is an important demonstration as currently this takes days for an experienced user to achieve this by hand on a single device and many (~1,000,000) devices require tuning in order to implement a technologically useful number of semiconductor spin qubits.
We have also developed an algorithm for the fine-tuning of semiconductor quantum devices (currently under review in New Journal of Physics). The algorithm performs a measurement and gives it a score. Gate voltage parameters are then adjusted in order to optimise this score. In this way, the algorithm proves capable of fine-tuning several important device characteristics at once. These characteristics would typically be fine-tuned by hand one at a time independently. Dominic was also involved in the development of the first reinforcement learning algorithm for quantum device measurement.

The main aim of the project was to develop an algorithm(s) capable of efficiently tuning up an electrostatically defined quantum dot for use as a singlet-triplet qubit. The above methods address the goals of the project in the following ways. The first algorithm aims to provide a way of efficiently characterising the state of a device, this in turn can be used to speed up the subsequent algorithms. The second algorithm demonstrates the course tuning of a device into a double dot regime faster than human experts. The formation of a double dot is required in order to operate a singlet-triplet qubit. Finally once a double dot regime is formed it cannot immediately be used, characteristics of the formed double quantum dot must be optimised. The final algorithm aims to optimise characteristics of the double quantum dot.
Exploitation Route We envisage that the research outcomes that were achieved being used in both academic and none academic settings to control, characterise, and develop semiconductor based quantum technologies. The developed algorithms are general allowing them to be applied to many different types of devices used in the field of semiconductor spin qubits. This means that the code we developed (available here https://github.com/oxquantum-repo) can be immediately be deployed and reused by others.
The algorithms are interesting for academics trying to develop new types of devices (in different host materials for example) as it allows them to automate and speed up the tune up and characterisation of new devices. It will also allow them to build larger and more complex devices containing more gate electrodes as these approaches are not limited to human tuning and characterisation methods.
For none academics they are also interesting as they provide methods for semiconductor device fabrication companies to improve their methods and benchmark performance.
Additionally some of the machine learning methods utilised are novel and hence they are of interest to machine learning researchers and technology companies.
Sectors Digital/Communication/Information Technologies (including Software),Electronics,Other

URL https://www.nature.com/articles/s41534-019-0193-4
 
Title Efficiently measuring a quantum device using machine learning 
Description The algorithm selects the most informative measurements to perform next by combining information theory with a probabilistic deep-generative model that can generate full-resolution reconstructions from scattered partial measurements. It is a general method capable of efficiently selecting the most informative point to measure next in a 2D map (image). 
Type Of Material Computer model/algorithm 
Year Produced 2019 
Provided To Others? Yes  
Impact We are aware of multiple groups who are interested in implementing this approach on various problems from microscopy to semiconductor devices. 
URL https://github.com/oxquantum-repo/CVAE_for_QE