Projects 2021

Hardware for combined flux and microwave control of transmon qubits

by Tanzan Araki, Lukas Mouton, Jing Huang, Luca Hofele

Host laboratory: Wallraff Group

Abstract: Scaling the number of qubits is one of the greatest challenges for realizing error-protected quantum computing. Within this QuanTech Workshop, we made first steps towards combining flux and microwave control of transmon qubits, under the motivation of improving scalability by allowing a smaller chip design with fewer ports required and fewer wires in general. For this, we investigated three different approaches to implementing a diplexer, and the best results were obtained using surface-mount capacitors and inductors. The diplexer PCBs are to be housed into mechanical enclosures and attached onto the base plate within a dilution refrigerator of the Quantum Device Lab. Finally, we constructed simulations in ANSYS to design the diplexed control line. Gauging the charge and flux couplings, the simulation matches the theoretical prediction for simple geometries, and can be the cornerstone for simulations of more intricate structures.

Identification of magic angle twisted bilayer graphene - an exploration of various methods

by Noah Kaufmann, Markus Niese, Lara Ostertag, Alexandra Mestre

Host laboratory: Ensslin Group
Abstract: Since the discovery of unconventional superconductivity in magic-angle twisted bilayer graphene in 2018, the interest in this graphene structure has gathered major interest. In the process of fabricating magic-angle twisted bilayer graphene stacks, high temperatures and mechanic disturbances in the stacking process can alter the designed twist angle of 1.1°. The yield of devices with correctly twisted samples is therefore low (around 15 %). A method that allows for early measurements of the twist angle, with the desired accuracy, will significantly increase the efficiency in the device fabrication. This QuanTech project explores three methods to distinguish twisted from untwisted samples at ambient conditions and discusses the integration of those methods into the fabrication process. The first approach we present is the measurement of the size of the Moiré pattern arising from the twist, with atomic force microscopy. However, the method did not work with the current sample fabrication process. Next, we show how the optical contrast of graphene under the microscope can be used to approximate the twist angle. Two procedures, that are ready to use, are described enabling the estimation of small angle with a systematic error of 0.3 degrees. Lastly, we explore how Raman spectroscopy gives insights into the phonon dispersion relation of the scanned sample and hence reveal information about the twist angle. While the intensity of the 2D peak (one of the main spectral features of the Raman spectrum of graphene) can be used to distinguish monolayer graphene from bilayer graphene, its shape allows us to determine whether a sample is twisted or not. Furthermore, the position of peaks that originate from the additional periodicity in twisted samples can be used to estimate the twist angle.
 

Development of a Multi-Client Digital Lock Box for Trapped Ion Experiments

by Giacomo Bisson, Max Glantschnig, Dominic Hagmann, Peter Tirler

Host laboratory: Home Group
Abstract: Experiments with trapped ions require lasers which are stable in frequency, amplitude and phase. To achieve this, one commonly uses active feedback, i.e. one measures the deviation of a quantity from its desired value and feeds this error back to the system. To tune this feedback loop, one commonly uses a PID (Proportional, Integral, Derivative) controller, which is also referred to as a lock box. The aim of this QuanTech project was to develop a replacement for the digital lock box EVIL that had been used at the Trapped Ion Quantum Information group over the past decade and which has become deprecated since. We evaluated different options and eventually opted for a semi-custom solution based on the commercial mixed signal Red Pitaya STEMlab 125-14 board. Special care was taken to ensure backward compatibility with the previous solution, both hardware and software wise. To this end, we designed a custom printed circuit board which we called Bichannel Lockbox On One Device (BLOOD) that can be deployed in the same electronics racks that are used for the EVIL, i.e. has the same form factor and power supplies. This PCB also comes with the possibility to change the gain and offset of the analog inputs and outputs digitally. It runs custom software, for which we combined the open-source PyRPL project, a software/firmware stack designed for controlling AMO experiments on the Red Pitaya with Python, with the so-called DEVIL, the software currently used to control the EVILs in the TIQI group. During the course of this project, we build a prototype of the BLOOD and we verified that the soft and hardware work together in the way we intended. However, to make it fully usable in the lab, some further developments are needed.
 

Quantum GANs for Quantum Process Tomography

by Gerard Aguilar, Elie Bermot, Sarah Hiestand, Jocelyn Terle

Host laboratory: Renner Group
Abstract: With the rise of quantum computing, the machine learning algorithms have been adapted to work on quantum devices. The main objective of this project was to assess the problem of unwanted scaling in complexity in the task of quantum process tomography by using quantum machine learning techniques, in particular qGANs. We selected the double-discriminator architecture, and showed successful results for learning a 1-qubit unitary process with the simplest ansatz possible and a variety of hyperparameters, all of which achieved convergence. For the 2-qubit case, we observed a much higher sensitivity to hyperparameter changes, something which we attribute to the large increase of the number of parameters to be optimized and new factors to take into account, like entanglement. While the approach could then be extended to larger number of qubits and also to learning CPTP maps, for those cases similar situations to the 2-qubit case arose, but we didn’t manage to find a good combination of hyperparameters that yielded convergence. In that respect, we strongly believe that a crucial next step would be to perform an in-depth study of convergence conditions of these models, in order to try to alleviate the amount of hyperparameter hand-tuning. While there are still some unknowns, such as how would the depth of the ansatzes need to scale up with the number of qubits –hence once again the importance of studying convergence and how this could affect it– we believe that this is an interesting approach that could potentially bring some quantum advantage and help some already existing machine learning protocols.

Investigating trainable hybrid quantum convolutional neural networks for image
recognition

by Odiel Hooybergs, Finn Janssens, Luca Righetti, Isidor Schoch
Host laboratory: Renner Group (https://qit.ethz.ch/)

Abstract: In the first part of this work we give an in-depth theoretical motivation for exploring trainability of quantum neural network performance on different applications. Next we investigate two different hybrid quantum-classical neural network architectures: one with a trainable fully connected quantum layer at the end and one with a trainable quantum convolutional layer in the beginning - where the trainability of this layer is a new step in the investigation of quantum convolutional neural networks. The actual implementation of these models is done in PYTHON 3, making use of the PENNYLANE and TENSORFLOW libraries. In this work the neural network is trained to classify handwritten digits of the MNIST database. We compare the results of the quantum neural network with a similar classical counterpart. Different encoding schemes and quantum circuits are investigated in the quantum layer. The model is able to train well, but for the small filter sizes used the trainable quantum circuit gets optimized to a circuit equivalent to the identity, resulting in no advantage over the classical convolutional neural network. A higher number of qubits (and thus the possibility to work with bigger filter sizes) and/or training on quantum data might result in a quantum advantage in terms of accuracy or speedup in the future.