SciPost logo

SciPost Submission Page

QArray: a GPU-accelerated constant capacitance model simulator for large quantum dot arrays

by Barnaby van Straaten, Joseph Hickie, Lucas Schorling, Jonas Schuff, Federico Fedele, Natalia Ares

Submission summary

Authors (as registered SciPost users): Jonas Schuff
Submission information
Preprint Link: scipost_202404_00010v2  (pdf)
Code repository: https://pypi.org/project/qarray
Date accepted: 2024-09-12
Date submitted: 2024-08-05 13:51
Submitted by: Schuff, Jonas
Submitted to: SciPost Physics Codebases
Ontological classification
Academic field: Physics
Specialties:
  • Condensed Matter Physics - Theory
  • Condensed Matter Physics - Computational
Approaches: Theoretical, Computational

Abstract

Semiconductor quantum dot arrays are a leading architecture for the development of quantum technologies. Over the years, the constant capacitance model has served as a fundamental framework for simulating, understanding, and navigating the charge stability diagrams of small quantum dot arrays. However, while the size of the arrays keeps growing, solving the constant capacitance model becomes computationally prohibitive. This paper presents an open-source software package able to compute a 100 x 100 pixels charge stability diagram of a 16-dot array in less than a second. Smaller arrays can be simulated in milliseconds - faster than they could be measured experimentally, enabling the creation of diverse datasets for training machine learning models and the creation of digital twins that can interface with quantum dot devices in real-time. Our software package implements its core functionalities in the systems programming language Rust and the high-performance numerical computing library JAX. The Rust implementation benefits from advanced optimisations and parallelisation, enabling the users to take full advantage of multi-core processors. The JAX implementation allows for GPU acceleration.

Author comments upon resubmission

We thank all the reviewers for their useful comments, which we think have helped us improve and strengthen our manuscript, code and documentation. We answer the questions of the reviewers here. Changes to the manuscript are reported separately and are written in blue in the updated manuscript version and marked with “>[…]<“ in the list of changes.

In response to questions from review 2
Question 1:Is the integer minimum unique? If the configurations of two dots is completely symmetric with respect to gate voltages, would the default algorithm and thresholded algorithm find both configurations or would they only find one?
Response: We have now clarified this point in the manuscript. Along a charge transition (shown as black lines in Figures 1 c and d), two charge states are indeed equal in energy. At T = 0 this constitutes an infinitesimally thin line. At these points, our algorithm would find only one of the change states due to the limits of numerical precision. However, at finite temperatures, the return charge state would be an equal mixture of the two.

Question 2:It is not clear to me that the thresholded algorithm always returns the correct integer number configuration. Is there an intuitive way to understand why the thresholded algorithm works for any potential? Fig 2b delivers the point, but the potential is fairly specific.
Response:We agree with the reviewer that, indeed, it is not true that the thresholded algorithm always returns the correct charge state(see Figures 7c, 8 and 9). As we discussed in the appendix, we observe artefacts particularly as distortions of the interdot charge transitions. Importantly, our results show that a correct choice of threshold can often reduce the simulation time without introducing errors. Based on comments from other reviewers, we have added a warning that advises the user to either increase the threshold or use the default algorithm if the value selected is found to be too low, based on eq. (F3). To further clarify this aspect, we have also added a new plot showing an example of error against the threshold value.
Regarding the reviewer’s second question, about the validity of the threshold algorithm for any potential, we claim that the class of potential such as those of figure 2b are indeed quite general, given the assumptions of the constant capacitance model. About this, we note that the prediction of the capacitance model is generally valid for weakly coupled quantum dots. For example, it is well known that double quantum dots that are interacting via a strong tunnel coupling exhibit clear deviations from the model predictions at the interdot charge transitions see (Ref. 1). For weak interdot coupling, assuming that the capacitive coupling between dots is weaker than the coupling between the dots and the gates, is also a valid assumption. As a result, the matrix $c_{dd}$ (and therefore $c^{−1}_{dd}$) must be approximately diagonal, leading to a fairly general class of potentials such as that of Fig 2b.

Question 3: Is it possible to combine quantum dots where some are electron dots and others are hole dots? The fact that Q⃗ = ±eN⃗ suggests not because elements N⃗ are defined as integers. Is this a limitation?
Response: The ability to combine both electrons and holes within the same array is not currently available in our software. This feature could be implemented in future versions of the software. We will work on this.

In response to questions from review 3:
Comment 1: Many functions in the code are not documented properly. Moreover, many docstrings are very short and hard to understand. Take the example of the function ”charge state contrast”. The docstring only says ”Function to compute the charge state contrast”. It is missing an explanation though what charge state contrast even is (google only finds 5 instances for this expression, so it is not widely used). I can guess it from the result, but I should not need to guess.
Response: We thank the reviewer for the valuable feedback. We have improved the doc-strings and documentation for the project and included a short description of each function’s functionality.

Comment 2: I don’t find the package very intuitive to use. I can run the examples, but I would find it hard to code my own simulation. A reason for this is that the example in the paper is very short and incomplete (not even the output is shown). The different concepts used by the classes that build the system are barely explained. There are many examples in the repository, but these also lack a proper tutorial-like documentation that would explain how to build the system, and how to use the different functions.
Response: We agree with the reviewer on the usefulness of a more tutorial like documentation and we apologise for not including this earlier. We have now added a ”read the docs” style documentation page (https://qarray.readthedocs.io/en/latest/index.html). This set of step-by-step explanations contains several examples that we hope will be useful to get new users ready to start coding their own models.

Comment 3: Moreover, for the advanced examples shown in Fig. 3, I found it hard to find the corresponding code in the repository: I eventually found the code for Fig. 3b by clicking on names that seemed relevant, but I gave up to find Fig. 3a after 5 minutes. It should not be that hard to find these examples.
Response: We agree with the reviewer and we apologise for this oversight. We have now renamed the code to reproduce these figures to fig_3b.py and fig_3d.py, respectively.

Comment 4: The package provides functionality to mimic experimental signatures, but I find this intransparent: The description in IV.A is very hard to follow. What is exactly done? Can this be explained with formulas/sketches? Moreover, Fig. 3b does some tricks to simulate the response to a QPC, but this is only explained in a single sentence in the caption
Response: We thank the reviewer for this comment. We have now clarified this in section IV.A and improved the description of the charge sensing simulation.

Comment 5: The authors argue for the validity of their algorithm in the appendix and show limits on the condition number of the capacitance matrix. They argue that this is usually fulfilled in experiments, but does the code actually check this criterion? Given that correctness is essential for the package, the default algorithm must be completely robust, and this check seems essential.
Response: We thank the reviewer for rising this important point. We have now added the suggested check in the revised version of the code.

Comment 6: I find it hard to judge the threshold algorithm. Is it possible to warn the user if the threshold is wrong, or is it up to the user completely? There are estimates for thresholds, but they are not good enough to be used automatically?
Response: We see the concern of the reviewer that the approximation made by the threshold algorithm may cause artefacts in the simulation. Following the suggestions also from other reviewers, in the revised version of our manuscript and code we have included a new figure and a warning to advise the user of the potential discrepancies and artefact that may arise using the thresholded algorithm. The new figure shows the discrepancy between a simulation with and without the threshold, as a function of the threshold value. As already mentioned in other reviewer’s comments, the newly introduced warning notifies the user when the selected threshold is below a value estimated from the capacitance matrix following Eq. (F3).

List of changes

Changes to the manuscript are written in blue in the updated manuscript version and marked with “>[…]<“ here.

In answer to review 1:

Requested change 1: Introduction : it would be interesting to have a few sentences in the introduction on the general background to the usefulness of charge stability diagrams.
Response: We thank the reviewer for this comment. As requested, we have modified the first paragraph of the manuscript.
Change in ‘I. INTRODUCTION’:
“Semiconductor quantum dot arrays are one of the leading platforms for the realisation of large-scale quantum circuits. As they grow in size and complexity, these arrays present exciting opportunities and significant challenges. >Tuning and operating these devices involves finding precise gate voltages to achieve the desired charge occupancy and interdot coupling parameters. Charge stability diagrams provide a visual representation of the charge occupation in the quantum dot as a function of the applied gate voltages, allowing the identification and control of transitions between different charge configurations.<
The constant capacitance model is a widely-used equivalent circuit framework that models the electrostatic characteristics of quantum dot arrays. >This model can be used to compute charge stability diagrams of quantum dots arrays [1-7]. As a result, the constant capacitance model< has proven to be an invaluable resource for gaining insights into complex charge stability diagrams and training neural networks for automated tuning strategies [2,5,8-11]. ”

Requested change 2: In Part II, it may be useful to restate the conditions for studying the system in the classical constant capacitance model approach.
Response: We have included the necessary condition for the constant capacitance model as per ref: [Rev. Mod. Phys. 79, 1217 (2007)] .
Change in “II. THE CONSTANT CAPACITANCE MODEL”:
“ The constant capacitance model describes an array of quantum dots and their associated electrostatic gates as nodes in a network of fixed capacitors [1–7], >(see Fig. 1a). The model is based on the assumption that a single constant capacitance C models the interactions between the charges on a single quantum dot and those in the environment and that the single-particle energy levels do not depend on the number of charges considered. Additionally, the model is a good approximation when the quantum dots are weakly coupled enough that quantum tunnelling effects can be ignored. Typically, the constant capacitance model is valid in a small range of gate voltages, as significant changes in the confinement potential can shift the dot positions and consequently alter their capacitive couplings. <”

Requested change 3: It seems to me that the formula for F(qd,Vg) should also be recalled before injecting equation (3) and obtaining (4).
Response: We thank the reviewer for this suggestion; we have updated the paragraph.
Change in “II. THE CONSTANT CAPACITANCE MODEL”:
“The free energy of the system, > $F(\vec{Q}; \vec{V}) = U(\vec{Q}; \vec{V}) - W(\vec{Q}; \vec{V}) = \vec{V}^T \vec{Q} / 2 - \vec{Q}_g \vec{V}_g$ < , can then be written as
\begin{equation}\tag{4}
F(\vec{Q}_d; \vec{V}_g) = \frac{1}{2}\vec{Q}_d^{T} c_{dd}^{-1} Q_d - (c_{dd}^{-1} c_{cd} \vec{V}_g)^{T} \vec{Q}_d,
\end{equation}”

Requested change 4: In part IV .B. equation (10) reintroduce Vg in the arguments of F.
Response: As requested, we have added the arguments to the function. Please note this is now equation (11) in the revised manuscript.
Change in “IV. ADDITIONAL FUNCTIONALITY”:
We have added the arguments to the function, which is now equation (11).

Requested change 5: On which configuration is the summation performed to obtain the soft- min in the case of thermal enlargement? It seems to me that it refers to all the configurations around the continuous optimum, otherwise the problem back to the brute-force case. This is not very clear in the paragraph.
Response: We agree with the reviewer, and we have now clarified this in the manuscript.
Change in “IV. ADDITIONAL FUNCTIONALITY”:
“The summation runs over all the allowed charge configurations >considered by the algorithm<. For a closed array, this means eliminating the charge states with the total number of charges different than $\hat{N}$. >This approximation is valid for temperatures such that $k_B T$ is smaller than the dot charging energies, since the contribution of other charge states is exponentially suppressed.<”

Requested change 6: I think it would be useful to add a few sentences on the benefits of charge sensors, optimal gates voltages and virtual gates. Perhaps a figure would be welcome here.
Response: We agree with the reviewer that highlighting the benefits of these functionalities will provide more context to the paper, and we have added a short paragraph.
Change in “IV. ADDITIONAL FUNCTIONALITY”:
“>These functionalities will be useful both when using QArray as a tool to understand charge stability diagrams and for automated tuning strategies. The charge sensing functions can be used to simulate realistic charge stability diagrams measured with a charge sensor. The optimal gate voltage function can be used to simulate a voltage sweep centred around a chosen charge state for any given capacitance matrix. Given a specific capacitance matrix, the optimal virtual gate function can be used to find the linear combination of gates that can modify the electrochemical potential of a single dot compensating for the effects of voltage cross-talk.<“

Requested change 7: Part VI: it seems to me that the article is rather allusive on the minimization method used (OSQP). I think this is an important point in the method (when the explicit result cannot be used) and the general principle should be explained. In addition, it doesn’t say how precise the continuous optimum must be, how many iterations must be performed, or how much computing time must be devoted to this step.
Response: We thank the reviewer for this insightful comment. We used OSQP since we found it was the highest-performance solver for the specific class of quadratic-constrained optimisation problems addressed in our manuscript. In the main manuscript, we have now included a new sentence that clarifies the choice of the solver and directs the interested reader to a more specialised reference (written by the solver’s authors).
Further on the reviewer comment: The number of performed iterations is determined by the solver such that the relative and absolute error in the solution falls below $10^{−3}$ (the default solver value). Given that our algorithms iterate over the nearest integer charge states, we find this precision acceptable.
Regarding the computational time, our benchmarks found the OSQP solver able to identify the continuous minimum of double, triple and even quadruple dot systems in less than 10 μs.
We have now clarified these aspects in the dedicated section of the manuscript (Section 3.A.2).
Change in “VI. IMPLEMENTATIONS”:
“>For a complete description of the OSQP solver and its benchmarks, see Ref. [27]. For the solver’s hyperparameters, we use the default values suggested by this reference. As a result, the solver performs the required number of iterations to obtain the continuous minimum to within an absolute and relative error of $10^{−3}$. The solver updates the step size automatically. Our benchmarks found that the OSQP solver was able to find the continuous minimum of double, triple and even quadruple dot systems in less than 10 μs per voltage configuration.<“

Requested change 8: the origin of the error bars in the figures and the method of obtaining them should be more explicit.
Response: We thank the reviewer for pointing this and we apologise for the oversight. We have now added a short description in the figure caption.
Change in “VII. BENCHMARKS”, in the caption of Figure 6:
“>The error bars represent confidence intervals of 2σ, such that for a randomly chosen set of capacitance matrices and voltages, the compute time should fall within the error bars 95% of the time. <”

Requested change 9: I am uncertain by Figure 5. Smaller values of t give a faster result. However, according to the authors, this is linked to a more approximate result. I think it would be good to be able to see on the figure the competition between calculation time and error committed than just running time.
Response: We agree with the reviewer that this would be rather informative and we have thus added the requested figure to the main manuscript (see new Fig. 7(c)). In addition, we wish to note that the thresholding strategy errors are further discussed in the appendix section J, see also Fig. 8 and 9. As specified in the paper, we noted that the threshold value after which significant discrepancies appear in the charge stability diagram depends on the parameters of the capacitance matrix. Hence, as further discussed in the appendix, a suitable choice of the threshold should be made by taking these parameters into account. To further clarify these points to the reader, we have added a new panel to (now) figure 7 of the main manuscript showing the percentage error introduced by the thresholded algorithm as a function of the chosen threshold. We estimated this error as a pixel comparison against the default (non-thresholded) version.
Finally, following also the advice of reviewer 3, we have coded explicit warnings for the user when the threshold is set too low based on equation (F3) of the appendix.

Requested change 10: Appendix C : Reintroduce Q notation, introduce L and lambda
Response: We have introduced the Q notation as requested. The references to L and λ were not meant to be there, so these have been removed.
Change in “Appendix C”:
New introduction to Q notation. The references to L and λ have been removed.

Requested change 11: the code lacks a complete description of all available functionalities and the general structure of the implementation (Doxygen type)
Response: Based on the feedback from all reviewers, we have substantially improved the documentation. There is an API overview of the whole package as part of this documentation.

Requested change 12: I find it confusing that some rust code is in .py and not .rs. In order to see which part of the code is in rust, python or jax, this could typically appear in the documentation.
Response: We assume the reviewer is referring to some thin Python wrapper functions present in the Qarray rust implementation folder. If so, we apologise for the confusion.
To clarify this point, all rust code is packaged as a separate Python package called “qarray- rust-core” (https://pypi.org/project/qarray-rust-core/). This package is a dependency of the QArray package. All code in this package is rust code (with .rs file names). The package is compiled using maturin (https://www.maturin.rs) so that it can be called by Python. The wrapper functions responsible for this confusion merely perform type conversions, such as f32 to f64, in the numpy arrays to be passed.

In answer to review 2:
Requested change 1:I would like to ask the authors to improve and extend the package tests.
Response: As requested by the reviewer, we have now expanded the testing suite such that 100% of the files and over 90% of code lines are covered by at least one test.

Requested change 2:I would like to ask the authors to document public functions and write tutorials, not just examples. The authors could also consider writing package documentation, e.g. using https: // readthedocs. org/ .
Response: Following the reviewer suggestion we have now written the readthedocs documentation, available at https://qarray.readthedocs.io/en/latest/introduction.html

Requested change 3:Not all figures are referenced in the main text, e.g Fig. 1a and 1b., and Section II would benefit from referring to Fig. 1a. Please ensure that all figures are referenced.
Response: We thank the reviewer for spotting this and we apologise for this oversight, we have now referenced all the figures in the main manuscript.
“Within QArray, the brute-force and our default algorithm are implemented in Python, Rust and JAX, >(see Fig. 1b)<.”
“The constant capacitance model describes an array of quantum dots and their associated electrostatic gates as nodes in a network of fixed capacitors [1–7], >see (Fig. 1a)<”.

Requested change 4: Gray and black lines in Fig. 1a are hard to distinguish. I would like to suggest using a different color or dashing one of the lines.
Response:We have modified the figure as requested.

Requested change 5: The paper contains a few typos (”Compted”), please fix these.
Response: We thank the reviewer for pointing this to us, we have fixed these typos in our revised manuscript.

Requested change 6: The references are not formatted, please format them.
Response: We have now reformatted the references.

In answer to review 3:
In answering comment 4, we have changed the following in IV.A:
“ To simulate charge-sensing measurements, we >distinguish between charge-sensing dots and regular dots. We first compute the charge state that minimises the free energy of the regular dots. With the charges on the regular dots fixed, we then determine the charge configuration on the sensor dots that minimises the overall free energy.
Next, we evaluate the difference in free energy between the minimum free energy charge state of the regular dots and the sensor's charge state next lowest in energy, which we denote as \(\Delta F(\vec{V}_g)\). We assume that the charge sensor dots are in the strongly coupled regime, so the conductance follows the Breit-Wigner formula [21]. Consequently, the charge sensor's response is Lorentzian, and we model the sensor response as:
\begin{equation}
s(\vec{V}_g) = \frac{1}{(\Delta F(\vec{V}_g) / \Gamma) ^2 + 1}
\end{equation}
where \(\Gamma\) is the sum of the tunnel rates to the source and drain; this parameter is configurable by the user to match their experimental setup. Additionally, we include various noise models to simulate white noise, \(1/f\) noise, and sensor switches due to fluctuating charge traps, among others.< ”

Requested change 1: Improve docstrings in code
Response: We have now updated the docstring to include descriptions of the functions, all their arguments and return values.

Requested change 2: Explain usage of code better (either in paper or in a separate documentation): add tutorials, explain the concepts to build the system better, show several explicit examples with extended explanation
Response: We have rewritten the example section to make it more user-friendly. In addition, on the readthedocs documentation page, we have written a set of examples covering:
1. Getting started
2. Using the GUI
3. Simulating change sensing measurements
4. Adding noise to the change sensing measurements
5. Adding latching to the change sensing measurements.

Requested change 3: Make code of all examples in the paper easy to find
Response: We thank the reviewer for highlighting this and we apologise for the oversight. The code examples are now clearly named figure 4b and figure 4d and are in the examples folder. https://github.com/b-vanstraaten/qarray/tree/main/examples

Requested change 4: Explain in detail how the experimental signatures (charge sensor dot) are calculated. Note that I found the term ”non-charge sensor dot” very confusing. I guess you mean the regular dots?
Response: We have made the requested change. Please see our reply on the reviewer comment 4.

Requested change 5: Eq. C.3 talks about differentiating L, but there is no L introduced in C.
Response: We thank the reviewer for noticing this, this was a typo. We apologise for the oversight and we have now fixed it.

Requested change 6: address the issues regarding validity mentioned in the report.
Response: We have made the requested changes, as per our answer to comment 5 of the reviewer we have now added these checks to the code.

Current status:
Accepted in target Journal

Editorial decision: For Journal SciPost Physics Codebases: Publish
(status: Editorial decision fixed and (if required) accepted by authors)

Login to report or comment