IEEE Transactions on

Quantum Computing uantum Engineering

Received 17 August 2023; revised 24 November 2023; accepted 11 December 2023; date of publication 15 December 2023;
date of current version 12 January 2024.
Digital Object Identifier 10.1109/TQE.2023.3343625

Quantum Vulnerability Analysis to Guide

Robust Quantum Computing System
Design
FANG QI1 , KAITLIN N. SMITH2 , TRAVIS LECOMPTE3 ,
NIAN-FENG TZENG4 (Life Fellow, IEEE),
XU YUAN5 (Senior Member, IEEE), FREDERIC T. CHONG6 (Fellow, IEEE),
AND LU PENG1 (Senior Member, IEEE)
1
Tulane University, New Orleans, LA 70118 USA
2
Northwestern University, Evanston, IL 60208 USA
3
Louisiana State University, Baton Rouge, LA 70803 USA
4
University of Louisiana at Lafayette, Lafayette, LA 70503 USA
5
University of Delaware, Newark, DE 19716 USA
6
University of Chicago, Chicago, IL 60637 USA
Corresponding author: Lu Peng (e-mail: [email protected]).
The work of Travis LeCompte was supported by a Louisiana Board of Regents Graduate Fellowship. This work was supported in part
by the National Science Foundation (NSF) under Grant OIA-2019511, in part by Enabling Practical-scale Quantum Computing
(EPiQC), an NSF Expedition in Computing, under Award CCF-1730449, in part by Software-Tailored Architecture for Quantum
(STAQ) under Award NSF Phy-1818914, in part by the US Department of Energy (DOE) Office of Advanced Scientific Computing
Research, Accelerated Research for Quantum Computing Program, in part by the NSF Quantum Leap Challenge Institute for Hybrid
Quantum Architectures and Networks under NSF Award 2016136, in part by the U.S. Department of Energy, Office of Science,
National Quantum Information Science Research Centers, in part by the Army Research Office under Grant W911NF-23-1-0077, and
in part by the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract
DE-AC05-00OR22725.

ABSTRACT While quantum computers provide exciting opportunities for information processing, they
currently suffer from noise during computation that is not fully understood. Incomplete noise models have
led to discrepancies between quantum program success rate (SR) estimates and actual machine outcomes.
For example, the estimated probability of success (ESP) is the state-of-the-art metric used to gauge quantum
program performance. The ESP suffers poor prediction since it fails to account for the unique combination of
circuit structure, quantum state, and quantum computer properties specific to each program execution. Thus,
an urgent need exists for a systematic approach that can elucidate various noise impacts and accurately
and robustly predict quantum computer success rates, emphasizing application and device scaling. In this
article, we propose quantum vulnerability analysis (QVA) to systematically quantify the error impact on
quantum applications and address the gap between current success rate (SR) estimators and real quantum
computer results. The QVA determines the cumulative quantum vulnerability (CQV) of the target quan-
tum computation, which quantifies the quantum error impact based on the entire algorithm applied to the
target quantum machine. By evaluating the CQV with well-known benchmarks on three 27-qubit quantum
computers, the CQV success estimation outperforms the estimated probability of success state-of-the-art
prediction technique by achieving on average six times less relative prediction error, with best cases at 30
times, for benchmarks with a real SR rate above 0.1%. Direct application of QVA has been provided that
helps researchers choose a promising compiling strategy at compile time.

INDEX TERMS Quantum computing, resilience, success rate (SR), vulnerability analysis.

I. INTRODUCTION problems. By carefully leveraging quantum superposition,

Excitement surrounds quantum computation due to the great interference, and entanglement, quantum computers are
theoretical potential both fault-tolerant [16], [44] and near- projected to be applied to computational tasks that are cur-
term [30] quantum computers have to solve high-impact rently intractable on today’s most powerful supercomputers.

© 2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 5, 2024 3100411
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 IEEE Transactions on
uantum Engineering Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN

quantifies the compiled circuit performance, and 1-CQV pro-

vides an SR estimation that closely models actual quantum
computer performance. By performing the QVA during com-
pilation, researchers will have a clear and direct view of how
gate error behaviors couple to the circuit structure to influ-
ence runtime performance. The blue box in Fig. 1 reveals the
brief structure of the QVA and how CQV would be used in
FIGURE 1. Quantum experiment workflow with QVA. One complete cycle an experimental flow to estimate the quantum computer SR
ranges from idea conception to real machine run. before real machine evaluation.
We extensively validated the accuracy and stability of
QVA in predicting the SR of any given compiled circuits.
Recent progress in quantum hardware has allowed many This assertion is backed by over 160 K experiments con-
prototype quantum computers to emerge, and superconduct- ducted on three state-of-the-art 27-qubit IBM quantum com-
ing circuits [5], [6], [45] are gaining popularity as one of puters [1], spanning six distinguished algorithms with vary-
the forefront quantum computing technologies. Compared to ing qubit sizes. The compiled circuits were produced using
other quantum hardware, superconducting quantum comput- a variety of compilation strategies and incorporated multiple
ers have advantages in scalability, microwave control, and error-mitigation techniques to address the diverse noise pro-
nanosecond-scale gate operation [7], [9], [10]. files encountered over months of experimentation. This com-
While promising, superconducting quantum architectures prehensive approach bolsters the extendibility of our analysis
are currently too error-prone to support programs targeted for to a broader range of circuits.
large-scale applications. Near-term superconducting quan- All results show that QVA maintains a stable SR estima-
tum computers suffer from various noise channels that de- tion via 1-CQV, providing on average 6x improvements (30x
grade both quantum information and computation. This noise in the best case) in relative prediction error over the widely
is difficult to fully characterize and causes retention and implemented ESP estimator. Additionally, we conduct a case
operational errors that vary both across-chip and between study to provide the quantum community with a direct appli-
quantum computers [46] significantly. Quantum error correc- cation of the presented method by choosing promising com-
tion was developed to accommodate occasional errors during piling strategies at compile time. Further, our QVA module
quantum computation, but current noisy intermediate-scale provides instructions for reconstructing the model based on
quantum (NISQ) era machines do not have the operator pre- a particular quantum device’s topology and error behavior. It
cision or device scale to implement this routine [40]. There- ensures its compatibility with various superconducting quan-
fore, NISQ quantum machines perform noisy operations as tum computers, irrespective of their vendor or technology.
errors can happen on any physical qubit at any time during Below are the contributions of our article.
program execution according to error rates characterized by
randomized benchmarking. Fig. 1 shows the full quantum 1) We design and build a lightweight error modeling
computing flow that transforms a research idea into an ex- scheme based on QVA and are the first that quantify er-
periment on a real quantum computer. ror rate impact propagating across the cnot gates with
Improving circuit success rate (SR) is a popular research the error rate reported by randomized benchmarking.
topic, but few studies examine the quantum computer and 2) We implement and evaluate a framework that calcu-
circuit-dependent errors that result in specific SR. A robust lates the unique CQV for an algorithm/machine pair-
noise model is essential for accurate SR prediction and devel- ing. We show that the proposed SR estimator, 1-CQV,
oping SR boost technologies such as error reduction, bypass, outperforms the current state-of-the-art SR estimation,
and compression. Unfortunately, existing methods like statis- ESP, by an average 6x less relative prediction error.
tical fault injection [15], [37], [41] and estimated probability 3) We highlight the scaling potential of CQV: as an algo-
of success (ESP) [47] have accuracy and scaling challenges. rithm reaches and surpasses the quantum volume of a
These models neglect the impact of circuit composition on device, CQV-based methods experience more than 10x
reported error rates from randomized benchmarking. There- improvement in relative prediction error rate compared
fore, this article proposes a systematic approach to explain with the state of the art.
different errors and provide accurate SR prediction.
While studying ESP, we found that there are gates whose II. BACKGROUND AND RELATED WORK
impact on the output is much more significant than their error A. NISQ ERA QUANTUM BASICS AND ERROR
rates. We propose the quantum vulnerability analysis (QVA) CHARACTERIZATION
to address these gate error inconsistent behaviors. QVA is a The flow for generating a quantum executable from a given
systematic methodology that performs error modeling based algorithm and evaluating it on a quantum chip is illustrated
on randomized benchmarking to determine the success of a in Fig. 1. The quantum compiler will be given information
compiled circuit on a targeted quantum computer: the vul- about the target quantum chip and compiler strategy, such as
nerability metric cumulative quantum vulnerability (CQV) optimization levels, initial layout method, mapping method,

3100411 VOLUME 5, 2024

 IEEE Transactions on
Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN uantum Engineering

etc. Based on that input, the compiler will follow all inter- B. RELATED WORK ON QUANTUM SR ESTIMATION
mediate compiling steps to generate a compiled circuit for Early quantum computing research was focused on designing
execution on a quantum computer. quantum hardware [28], instruction set architecture [8], and
When operating a superconducting quantum computer, quantum computer microarchitecture [9], [10], [32], [33].
operations may fail due to poor environmental conditions, Afterward, the temporal and spatial noise variation chal-
inaccurate control, state decoherence, and more. The error lenges of SC quantum computers were studied to discover
rate associated with an operation, operational error rates, is mapping and allocation-enhanced compilations to make al-
closely estimated by randomized benchmarking, described gorithm execution more robust to diverse errors [20], [21],
in detail below, which approximates the extent of failures [22], [31], [46]. Additionally, works such as [18] contribute
without revealing the exact source. The lifetime of a qubit, to the understanding of how noise, fidelity, and computa-
or its ability to retain a quantum state, is determined by its tional cost interplay in quantum processing, enriching the
relaxation time (T1) and decoherence time (T2), so-called broader discourse on quantum system performance. Cur-
retention errors. The decoherence and relaxation time rep- rently, the focus of quantum computing is on optimizing the
resent the qubit’s average time to retain its energized and SR by applying different compiler strategies, such as miti-
superimposed states, respectively. gating the effect of errors by enhancing the quantum instruc-
Randomized Benchmarking: Operator performance must tions [14], [43], decreasing measurement errors [13], [48],
be accurately characterized to use a quantum computer ef- mitigating crosstalk errors [7], [34], combining preexecution
fectively. Unfortunately, quantum computer noise models and postexecution software approaches to improve perfor-
are complex, and it is unscalable to completely character- mance [19], [47], and compiling with specific constraints [7],
ize system noise via process tomography [39]. In addition [23].
to scaling considerations, characterization procedures must While many studies have improved quantum program per-
separate noise associated with quantum gates from errors formance, two areas have been underexplored: 1) accurate
stemming from state preparation and measurement to ensure SR prediction for specific compiled circuits and 2) better
that computation quality can be adequately estimated. Ran- modeling of error/algorithm relationships in current quantum
domized benchmarking [26], [27] is a method of assessing systems. Regarding SR prediction methods, popular alterna-
quantum computer hardware that achieves an average error tives to noisy quantum computer simulations include using
rate for operations through a process known as twirling. At machine learning for SR prediction, which treats the entire
the high level, twirling implements long sequences of ran- computation as a black box [25], developing detailed noise
dom gate operations and fits the resulting data to a curve models, and methods like statistical fault injection [17], [37],
to determine the average error. Because randomized bench- [41] and estimated success probability (ESP) [36], [46], [47]
marking considers only the exponential decay of sequences
of random gates, sensitivity to measurement noise in the gsi = (1 − gei ) msi = (1 − mei ) (2)
resulting average error is minimized. Meanwhile, the T1 and Ngates
T2 errors on gate operation are included as part of the gate’s  
NMeasurement
ESP = gsi ∗ msi . (3)
error rate reported by the randomized benchmarking [36].
i=1 i=1
Randomized benchmarking, while applicable to systems of
any dimension, is predominantly employed for single-qubit
III. MOTIVATION
or two-qubit gates [11]. This preference arises from the expo-
A. LIMITATION OF CURRENT SR ESTIMATOR
nential growth in the number of required gates as the system’s
1) MACHINE LEARNING BASED
dimension increases, making the method less practical for
larger systems. Despite this, the technique can be adapted The machine learning-based success rate (SR) prediction
to pinpoint errors due to unintended crosstalk [12]. No- method, referenced in [25], simplifies quantum computations
tably, IBM utilizes randomized benchmarking in each cali- into a black box model. This approach can blur distinctions
bration cycle to ascertain error rates for their quantum com- between circuits with varied gate parameters and requires
puter’s single and two-qubit gates, as reflected in the system’s retraining when adapting to different quantum machine sizes.
properties Data collection for larger machines is resource-intensive, es-
pecially given the method’s requirement to gather data within
a single calibration period. Furthermore, its validation, lim-
Trial counts o f correct out put ited to specific compiled circuits, may not account for the
Success Rate(SR) = (1)
Total trail counts. complexities of larger machines or diverse error-mitigation
strategies.
Success Rate: The SR is used to gauge quantum program
performance on a quantum computer. We compute SR by 2) FAULT INJECTION-BASED
dividing the number of correct outputs by total executions, The statistical fault injection method employs classical com-
as shown in (1). For more details on quantum computing, we puters to simulate full-state quantum computations. During
refer to [35]. this process, errors are systematically injected into each basis

VOLUME 5, 2024 3100411

 IEEE Transactions on
uantum Engineering Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN

gate based on specific triggering probabilities [15]. A very

recent work [37] employs fault injection methods to evalu-
ate quantum vulnerabilities. This study proposes the use of
quantum fault injection to scrutinize circuit vulnerabilities,
with a specific emphasis on radiation-induced errors. We
note, however, that the fault injection approach may face
challenges when applied to large machines. This complex-
ity is accentuated when sampling a broad spectrum of cir-
cuits that exhibit variations in qubit size, circuit depth, and
concurrent fault counts, especially in the context of larger FIGURE 2. Current success rate estimators performance.
machines and intricate algorithms. Moreover, the quantum
vulnerability factor (QVF) metric proposed in [37], can face
challenges when assessing circuits reaching the quantum
supremacy, typically seen in machines with 50+ qubits [4].
The QVF calculation hinges on the contrast function, which
necessitates prior knowledge of P(A), the expected correct FIGURE 3. ESP for different circuits.
state. In the absence of this knowledge, the correct state must
be determined through a noise-free simulation. Relying on
classical computing for full-state quantum circuit simulation its relatively shallow depth, allowing us to demonstrate the
poses a significant challenge, as such methods approach the effects of significant algorithm size increments. Conversely,
limits of classical computational capabilities. Alternatively, the QFT, characterized by deeper circuits, exhibits only mod-
our research introduces a different approach, formulating a est size increases. For a comprehensive evaluation, both algo-
noise model addressing 1- and 2-qubit gate errors, measured rithms were tested with three different input sizes, yielding
errors, and crosstalk errors. Our methodology is crafted with actual SRs ranging from 70% to 5%. ESP_CP, a variant of
a focus on scalability. ESP, focuses solely on multiplying the SRs of gates situated
on the critical path. Unfortunately, the predicted outcomes
3) ESTIMATED SUCCESS PROBABILITY-BASED from both methods show a significant deviation from the
The estimated success probability (ESP), shown in (3), pre- actual SR, with discrepancies ranging from 25% to 60% and
dicts the correct output trial probability by multiplying the relative error rates ranging between 70% and 470%. Further
SR, or fidelity, of each gate (gsi ) and measurement (msi ) op- compounding the issue, both methods produce SR estima-
erations, generated by one minus the gate (gei ) and measure- tions that diverge sharply from the real machine results as the
ment (mei ) error rate in (2). While ESP considers all circuit circuits increase in size, indicating that their tendency toward
operations, the product treats all gate errors that contribute scaling is not well performed.
to the final SR estimation equally when some gate errors
influence the final circuit outcome more or less than others. B. SOME ERRORS MATTER, WHILE OTHERS DO NOT
As a basic demonstration of the inaccuracy of ESP modeling, The ESP model, upon closer examination, presents poten-
the gate success products in (3) commute, whereas most op- tial sources of inaccuracies in SR prediction. Illustrated in
erations in quantum circuits are fixed in ordering [29]. Based Fig. 3, the ESP model accurately predicts the SR only for
on such position differences of the gates on the compiled the middle circuit. When considering a compiled circuit, its
circuits, gate errors will contribute differently based on their final SR results from the product of individual qubit SRs,
propagation path to the measurement, which influence the which are premeasured. Therefore, only errors impacting a
estimated SR differently than their original gate error rates. measurement gate influence the final output.
The detailed analysis is presented in Section III-B. On the For scenarios akin to the first circuit, the ESP model tends
other hand, the simplicity of the ESP metric has made it to overestimate error rates. Here, the error from the red-
frequently applied in quantum compiler design and circuit boxed Z gate does not influence the measurement gate, mean-
optimization efforts as a method to predict quantum program ing the Z gate’s error only impacts the SR of Q2. Yet, this is
success on quantum computers [2], [3], [24], [32], [36], [38], not captured by the subsequent Q1 measurement gate.
[47]. However, if a better SR estimator was available, the Contrastingly, for situations resembling the third circuit,
effectiveness of the aforementioned quantum computer opti- the ESP model tends to underestimate error rates. Errors orig-
mizations could potentially experience significant improve- inating from the two red-boxed H gates not only affect the
ments. measurements of their associated qubits but also influence
We used statistical fault injection, ESP_CP [46], and ESP, other qubits via the cnot gate. Despite this, the gate error
as illustrated in Fig. 2, to estimate the success rate (SR) to SR transformation, as outlined in (2), only accounts for
of the Bernstein-Vazirani (BV), and quantum Fourier trans- this error once. Any subsequent error impacts on other qubit
form (QFT) algorithms across varying scales on three distinct measurements arising from error propagation are disregarded
quantum computers. The BV algorithm was selected due to in the ESP model.

3100411 VOLUME 5, 2024

 IEEE Transactions on
Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN uantum Engineering

B. CQV DETERMINATION
1) CIRCUIT ARRAY GENERATOR
To understand the error propagation path within a quantum
circuit during runtime, a connection between when an error
occurs and how much it affects the compiled circuit must
be established. For more granularity, we quantify the com-
piled circuit to a finer degree by representing the algorithm
at the cycle level. Cycle-level representation for a quantum
circuit is analogous to the classical electrical circuit diagram
to replace the previous analysis at the level of the complete
compiled circuit. The circuit array generator block transfers
the compiled circuit further to a 2-D array where each el-
ement represents the attributes of the physical qubit at that
cycle, as shown in Fig. 4(b). The circuit array records each
physical qubit’s attribution in every cycle, including the gate
type, gate error, associated virtual qubit, its cumulative SR,
etc. Based on the cycle level compiled circuit, a snapshot of
the operating quantum chip at a given cycle can be linked
with the corresponding cycle in the compiled circuit.

2) CALCULATING (1 − CQV )
The CQV methodology aims to predict the failure rate for
a given compiled circuit on a designated quantum chip by
effectively modeling the errors based on its propagation. The
FIGURE 4. QVA workflow and circuit array example. Calculating 1 − CQV block of Fig. 4(a) will first receive
a circuit array with attributes filled. Next, we perform a
crosstalk error calibration based on [34] and update it to the
Such observations underscore that certain errors exert gate error by multiplying their SR based on (2). To deter-
more influence on the final output than others, highlighting mine the SR estimation, which is 1 − CQV , we introduce
the need for a refined approach. This analysis emphasizes an algorithm (referenced as Algorithm 1). This algorithm
the importance of taking into account circuit structure and progressively updates the cumulative success rate (CSR) of
architectural vulnerabilities when estimating SR on actual each physical qubit based on the circuit array, considering
quantum computers. the propagation of errors.
The algorithm initiates by setting the CSR for all entries
in the circuit array to a perfect score, i.e., CSR = 1 (100%
IV. QUANTUM VULNERABILITY ANALYSIS success), as depicted in line 3. Subsequent steps, from lines
A. QVA OVERVIEW 3 to 16, loop through all the gates, updating CSR values. For
In Section III-B, we discover that the SR of quantum compu- single-qubit gates, lines 6 and 7 modify the CSR for that gate
tation is the outcome of the SR of each qubit being measured. by multiplying its total SR with the preceding CSR value of
Each measured qubit’s correctness is influenced by gate er- the same qubit from the last cycle.
rors propagated to it. Our proposed QVA is a systematic Complex operations, like the cnot gate, necessitate a
methodology that follows error propagation. The QVA will deeper understanding. Here, errors from one qubit can cas-
estimate the vulnerability of the compiled circuit by perform- cade to itself and affect the paired qubit. In lines 8–11, for ev-
ing error modeling based on the error rates from randomized ery occurrence of such two-qubit interactions in a given com-
benchmarking calibration. piled circuit, we introduce a weight w. This weight, which
The QVA generates the cumulative quantum vulnerability lies between 0 and 1, signifies the fraction of cumulative error
(CQV) metric. Definition of CQV: CQV presents the final originating from the paired qubit that might propagate via the
circuit’s vulnerability by predicting the failure rate (FR) for cnot gate.
the compiled circuit. We emphasize that the CQV will not This error propagation model stems from the constraints
predict the correct result but the possibility of an incorrect imposed by the error rates disclosed through randomized
result during runtime. The calculation of 1 − CQV represents benchmarking. The intricacy lies in the fact that these re-
the estimated SR of a compiled circuit on the target machine ported error rates cannot be disaggregated into individual
calculated with the CQV. In Fig. 4(a), we present the com- types, such as phase, bit, or decoherence errors. Each type
plete workflow of QVA. behaves differently when channeled through the cnot gate.

VOLUME 5, 2024 3100411

 IEEE Transactions on
uantum Engineering Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN

Algorithm 1: Calculate (1 − CQV ).

Input:Physical qubits QP;Compiled Circuit Cycles C;
Weight w; Circuit Array CA = [QP][C + 1][Attr.];
Output:(1 − CQV )
1: (1 − CQV ) = 1
2: let Attr.qp,c = CA[qp][c][Attr.] for all Attributes
3: Initialize CSRqp,0 = 1.00 for every qubit at first cycle.
4: for each cycle c from 1 to C + 1 do
5: for each physical qubit qp in QP do FIGURE 5. Average success rate prediction comparison between weight
6: if the gateqp,c is 1-qubit gate then as 0 and 1 for benchmarks on IBMQ_Montreal.
7: CSRqp,c = gsqp,c ∗ CSRqp,c−1
8: else if the gateqp,c is CNOT gate then
9: crosserror = (1 − CSRqp ,c ) ∗ w
different error profiles. Then we used a machine learning-
10: CSRqp,c = gsqp,c ∗ CSRqp,c−1
based method to learn the weight value and used it to infer
∗ (1 − crosserror)
the proper weight in the CQV calculation. For more details.
11: end if
Please refer to Section V.
12: end for
13: if any qp in the final swap cycle then
14: swap the CSRqp,c and CSRqp ,c for all swap pairs
V. DETERMINING CNOT WEIGHT
15: end if
To accurately predict the real SR, QVA requires a proper
16: end for  weight value, between zero and one, at compile time to assist
17: 1 − CQV = CSRqp,c
with the CQV calculation. Before identifying the value of
gateqp,c =Measure
the weight, we first demonstrate the prediction performance
18: return 1 − CQV
when the weight is set equal to zero or one representing
no error or full error crossover the cnot gate, respectively.
For a target qubit involved in a cnot gate, its CSR is The compiled circuits of the experiments are generated from
computed as a product of its own SR (gs ), its preceding CSR, different combinations of compiler settings for four bench-
and the SR inherited from its paired qubit. This inherited SR marks at two different algorithm sizes. The CQV calculation
factors in only the weighted portion of the cumulative error. is performed using the calibration error and execution results
It is crucial to remember that the SR (or the associated error from IBMQ_Montreal on April 1, 2022. As shown in Fig. 5,
rate) for any quantum element (qubit, gate, or circuit) can though the CQV results with weight set to zero, 1 − CQV0 ,
be deduced using (2), which is based on the complementary are closer to the real SR than ESP, there is still a nontrivial
relationship between success and error rates. gap between 1 − CQV0 and the real SR meaning that some
For idle cycles in the quantum circuit, we attribute an error errors are not well represented. Meanwhile, 1 − CQV1 sets
value of zero. In the case of repeated gates, either compiler- the weight to one, which makes the predictions close to zero
introduced or manually inserted by the programmer, we ap- all the time and lose track of the real SR, meaning the errors
point the error value based on the known error rates of such are being overestimated. The experimental results show that,
gates in the target machine. for those benchmarks, using zero or one as the weight will
In conclusion, the 1 − CQV value, which represents the lead to inaccurate predictions. When brute-force performing
overall SR prediction, is derived by multiplying the CSRs the CQV prediction for all the weights with 1% granularity,
of all measurement gates. Upon examining Algorithm 1, we found the correlation between the weight value and its
it becomes evident that its complexity is O(G), where G corresponding 1-CQV prediction is approximately −1. In
represents total gate counts for all types. This complexity other words, among all the weight values, there will always
arises because the algorithm iteratively checks each physi- be one and only one weight value that returns a SR prediction
cal qubit’s gate in every cycle to update the corresponding closest to the real SR, which will be labeled as the best
cumulative SR. Consequently, the algorithm scales linearly weight.
with the gate count, encompassing all gate types, including Based on such observation, we calculated the best weight
ideal gates. This linear scalability of our noise model offers a for all compiled circuits generated from combining all the
significant advancement, facilitating both current NISQ-era different algorithms, target machines, and compiled strate-
and future quantum computing endeavors without being con- gies. As shown in Fig. 6, we have plotted the best weight
strained by the limitations of classical computational power. against the cnot count of the compiled circuit. The result is
consistent with our expectation—the best weight will be very
3) CNOT WEIGHT SELECTION arbitrary when the cnot count is low, but as the cnot count
In Section V, we have provided a study to observe the of the compiled circuit increases, the best weight begins to
weight selection impact for various compiled circuits with approach zero and shows an overall decreasing trend.

3100411 VOLUME 5, 2024

 IEEE Transactions on
Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN uantum Engineering

TABLE 1. Benchmarks and Quantum Computer Description

FIGURE 6. Comparing the best weight among all experiments.

VI. IMPLEMENTATION
We employ Qiskit [15], a renowned open-source framework
for quantum computing, as the foundation for implement-
ing and assessing our QVA methodology. Our work extends
Qiskit version 0.34.2, enabling it to execute QVA and com-
pute the (1-CQV) for any specified compiled circuit. Our
FIGURE 7. Graphic neural network-based model layout. QVA approach is meticulously crafted to provide an accurate
and efficient estimation of the SR for compiled circuits on
specific quantum machines. To ensure a diverse set of com-
piled circuits, we utilize various combinations of compiling
policies for each benchmark-machine pairing. This approach
integrates a spectrum of error-mitigation techniques at ev-
ery stage, ensuring comprehensive and robust evaluations.
Table 1 describes the backends and benchmarks used in our
evaluation. We repeatedly perform all the algorithms that
range in scale from 4 input qubits to 15 input qubits over
months, which generated 160 K distinctive compiled circuits
captured with diverse noise calibration profiles. To focus on
meaningful results, we ignore experiments that return a real
SR below 0.1%. We chose the state-of-the-art SR estimator
ESP for the baseline and ignored the ESP_CP since it under-
estimates error.
In the experimental flow, we first run the given compiled
circuits on their target quantum computers for the default
FIGURE 8. GNN training performance for IBMQ_Mumbai. 8192 trials and log the outputs. Then, we use the Qiskit
simulator to capture the correct result and generate the SR of
the experiment based on the logged output on the classical
After analysis, we found that many factors, such as ma- computer. For scenarios approaching quantum supremacy,
chine error properties, compiled circuit properties, etc., in- where classical computers struggle with full-state quantum
fluence the best weight value. To fulfill the need of taking circuit simulations, we introduce an alternative method to
the graph-like machine information and circuit features into ascertain correct results, as detailed in Section VII-B. Now
consideration, we choose a graph neural network (GNN) [42] the experiment’s real performance is known and named real
and combine it with feed-forward networks to perform the SR. For each machine, we used the first ten days of data to
best weight prediction shown in Fig. 7. The node matrix perform the GNN training on weight selection and to infer
represents the information for every physical qubit, including the weight to assist CQV prediction for the rest of the ex-
single-qubit operation error rates. The edge matrix in the periment data. This offline training is a one-time procedure,
figure presents the cnot error rates for each physical qubit and in our experiments, it took approximately an hour on
pair. The circuit matrix contains information for each qubit’s an NVIDIA 3060 GPU. Then we perform ESP and 1-CQV,
operation count, measurement info, and two-qubit operations which estimate the real SR based on the calibrated error of
count. As shown in Fig. 8, by leveraging the GNN model, we the experiment on the classical computer. The estimated SR
can provide the weight value with an average 2% difference generated from ESP and 1-CQV will be compared with the
from the best weight, which strongly supports an accurate SR real SR. In addition to quantifying the estimation accuracy
prediction in later execution. The trained model will be used directly by performing the absolute difference between the
to infer the best weight for CQV calculation. real and estimated SR, we also use the relative prediction

VOLUME 5, 2024 3100411

 IEEE Transactions on
uantum Engineering Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN

FIGURE 9. Predicted and real success rates of QPE with IBMQ _ Montreal Quantum machine. For each compile configuration listed on the x-axis, the
variables separated by an underscore are ordered as algorithm input qubits, optimization level, allocation method, and routing method.

FIGURE 10. Average relative predict error for all the benchmarks on FIGURE 11. Average relative predict error across all benchmarks and
single Quantum machine IBMQ_Montreal. backends.

error metric, which uses the absolute difference divided by prediction error by the SR. The trend emerges that 1-CQV
the real SR, to present the accuracy trend of the method while prediction outperforms the ESP by achieving an average of
compensating for increases in the algorithm size resulting in six times less relative prediction error rate and the best im-
a decrease in real SR. provement of 30 times. Meanwhile, the relative prediction
error jumps clearly when the algorithm size increases, which
VII. RESULTS follows the nature of the benchmark by employing signif-
A. CQV ACCURACY icantly more two-qubit gates. When increasing algorithm
1) ALGORITHM SIZE WITHIN QUANTUM VOLUME size, not only does the number of two-qubit gates increase,
Here, we present the CQV prediction performance for all but the number of swapping operations also increases. As
six benchmarks on the Quantum machines IBMQ_Montreal, shown in Fig. 11, 1-CQV presents a stable and accurate
IBMQ_Toronto, and IBMQ_Mumbai. As shown in Fig. 9, average relative prediction error across all machines and
we presented the 1-CQV prediction accuracy compared with benchmarks. Based on the results, we conclude that the CQV
ESP with varying compiled circuits for a five-qubit QPE al- achieved the goal of designing a more precise SR estimator
gorithm on IBMQ_Montreal on April 5, 2022. The different consistently across different dates, machines, and algorithms
configurations are guidelines for the compiler to generate the than the state-of-the-art SR estimator.
final compiled circuit based on the given logical circuit and
target device. 2) ALGORITHM SIZE BEYOND QUANTUM VOLUME
We observe that 1 − CQV is much closer to the real SR From the results shown in Figs. 9–11, 1 − CQV proves to
than the ESP. After being shown to the right of Fig. 9, the predict SR with a closer distance to the real SR, even when it
average absolute error rate for ESP over all the configura- falls in the 10%–0.1% range. The primary reason for the low
tions compared to ground truth SR is 29.8%, while 1 − CQV SR is that the size of the compiled circuits equals or exceeds
achieves an average error of 4.8%. Such an error rate differ- the desired quantum volume of the target quantum machine.
ence means that 1 − CQV achieves an 84% error reduction Quantum volume can be defined as the product of the number
compared to the ESP, which also means the CQV calculation of virtual qubits and maximum circuit depth supported by the
is adaptable to the variation of errors across different calibra- machine.
tion periods and provides excellent predictions. After making a full-spectrum comparison among all the
As shown in Fig. 10, we present the 1-CQV prediction for backends and benchmarks, from the observations of the re-
all the benchmarks on the single machine IBMQ_Montreal. sults, we can say that the CQV has better error modeling than
We include experiments with an SR higher than 0.1%. The the ESP noise model. Fig. 10 demonstrates that the CQV
relative prediction error is produced by dividing the absolute prediction is stable across different algorithms with a much

3100411 VOLUME 5, 2024

 IEEE Transactions on
Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN uantum Engineering

TABLE 2. Execution Time Comparison (Seconds)

lower relative error rate. Additionally, the results show that

CQV performs much better when the algorithm reaches or FIGURE 12. Choose compiler policies for HS on IBMQ_Toronto.
exceeds the quantum volume, which is a valuable property
when the limited quantum volume is the bottleneck of the
current NISQ era. Therefore, we conclude that the 1 − CQV performance and fine-tune our noise model. It is pertinent
prediction is accurate across the full spectrum of algorithm to note potential overfitting for specific circuits due to the
sizes and SRs. reliance on reverse circuits. Nevertheless, we believe our ap-
proach adds a valuable technique to the toolbox for exploring
B. SCALABILITY ANALYSIS OF CQV circuit vulnerabilities.
In the rapidly evolving landscape of quantum computing, Furthermore, while our preliminary results on QVA’s scal-
the intricacy of quantum circuits is escalating, bringing us ability are promising, they represent just the tip of the
closer to addressing real-world challenges. As we navigate iceberg. Our model’s design is inherently versatile, unen-
this frontier, the need for prediction models that are both cumbered by specific error rates, gate types, or circuit struc-
accurate and scalable becomes paramount. Our study, there- tures. This adaptability not only allows for the integration of
fore, focuses on the scalability of the QVA, examining its partial error correction techniques in larger devices but also
efficiency across a spectrum of quantum circuit sizes. hints at the vast potential for future optimization strategies,
The QFT was chosen as the benchmark for our evaluation. further refining the SR estimation process.
With input sizes spanning from 5 to 120 qubits, it is note-
worthy that machines operating with around 50 qubits are VIII. CASE STUDY: CHOOSING COMPILING STRATEGY
on the threshold of quantum supremacy [4]. Our scalability The current access modes for quantum computers are either
experiments for QVA were conducted based on the topology limited free access to small machines or expensive hourly
and error profile of the IBMQ_Washington machine, a state- institutional subscriptions to large devices. Naturally, users
of-the-art 127-qubit quantum computer. will want the highest SR with as few executions as possible
To ensure our execution time assessment was both trans- to save time, money, and access. However, finding the best
parent and unbiased, we considered only the inference time combination among all the available machines and compiler
of the GNN, which provides the cnot weight value, and configurations to achieve the best performance is challeng-
the execution time of the CQV noise model. This choice ing. The search space will grow enormously when also con-
was made because the GNN training is an offline process, sidering compilation optimizations. Without performing a
executed just once. As shown in Table 2, our CQV approach brute-force execution of all the combinations, identifying the
demonstrated linear scalability, with execution time increas- best strategy is challenging. Since CQV is more accurate than
ing proportionally with the cnot gate counts. It is worth ESP, we would naturally ask, could CQV be used to suggest
mentioning that the execution of the CQV takes place on the compiling configuration with the optimal performance?
the CPU, including the pretrained GNN to infer the weight To answer that, we performed both ESP and CQV for all the
value w. This linear trajectory underscores the potential of combinations of compiler configuration at compile time for
our model to efficiently manage complex quantum circuits the HS benchmark and picked the two configurations with the
in the future. highest estimated SR for both ESP and CQV. Based on the
However, the full-state quantum circuit simulator-based result, the CQV outperforms the ESP across all algorithm
approaches face challenges in scalability, particularly with sizes. One example of HS at level 2 optimization and six
circuits that exceed 32 qubits. This limitation highlights the input states is shown in Fig. 12. It is clear that the two choices
simulator’s constraints when tasked with emulating large- with the highest 1-CQV estimation not only have less predic-
scale, real-world quantum systems. In contrast, our approach tion error but also result in the highest real SR. In contrast,
requires only a fraction of the computational power for esti- the top two high-ranking ESP configurations result in the 4th
mating SRs. To both validate our predictions and refine our and 5th best in real SR out of nine combinations. Moreover, it
noise model, particularly for circuits approaching quantum will be the 7th and 8th choice of ESP to pick up the compiler
supremacy, we have devised a novel method: by merging the configurations for the highest real SR. For the computation
original circuit with its inverse and using the input states overhead, executing CQV prediction is acceptable since the
as a benchmark, we can effectively measure the circuit’s calculation is done on a classical computer. Furthermore,

VOLUME 5, 2024 3100411

 IEEE Transactions on
uantum Engineering Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN

since the execution of all the SR predictions is independent [7] Y. Ding, P. Gokhale, S. F. Lin, R. Rines, T. Propson, and F. T. Chong, “Sys-
of each other, it is possible to perform parallel computing for tematic crosstalk mitigation for superconducting qubits via frequency-
aware compilation,” in Proc. 53rd Annu. IEEE/ACM Int. Symp. Microar-
different compiler strategies, and the individual prediction chitecture, 2020, pp. 201–214, doi: 10.1109/MICRO50266.2020.00028.
overhead is discussed in Section VII-B. In this case, the CQV [8] X. Fu et al., “eQASM: An executable quantum instruction set architec-
can guide a user to choose a more effective and reliable com- ture,” in Proc. IEEE Int. Symp. High Perform. Comput. Archit., 2019,
pp. 224–237, doi: 10.1109/HPCA.2019.00040.
piler strategy with a higher SR than ESP. No prediction can [9] X. Fu et al., “A microarchitecture for a superconducting quantum
be perfect (that would be computationally intractable), but processor,” IEEE Micro, vol. 38, no. 3, pp. 40–47, May/Jun. 2018,
CQV improves prediction enough to be usable for compiler doi: 10.1109/MM.2018.032271060.
decisions. [10] X. Fu et al., “An experimental microarchitecture for a superconducting
quantum processor,” in Proc. 50th Annu. IEEE/ACM Int. Symp. Microar-
chitecture, 2017, pp. 813–825, doi: 10.1145/3123939.3123952.
[11] J. P. Gaebler et al., “Randomized benchmarking of multiqubit gates,” Phys.
IX. CONCLUSION Rev. Lett., vol. 108, no. 26, 2012, Art. no. 260503, doi: 10.1103/Phys-
In the rapidly evolving field of quantum computing, predict- RevLett.108.260503.
ing the SR of a quantum circuit remains a challenging task. [12] J. M. Gambetta et al., “Characterization of addressability by simultane-
ous randomized benchmarking,” Phys. Rev. Lett., vol. 109, no. 24, 2012,
Existing methodologies often fall short, either by oversim- Art. no. 240504, doi: 10.1103/PhysRevLett.109.240504.
plifying the error model or by not adequately accounting for [13] P. Gokhale et al., “Optimization of simultaneous measurement
the intricacies of error propagation within complex quantum for variational quantum eigensolver applications,” in Proc. IEEE
circuits. Recognizing this gap, we present the QVA in this Int. Conf. Quantum Comput. Eng., 2020, pp. 379–390, doi:
10.1109/QCE49297.2020.00054.
article, a robust systematic approach to determining a given [14] P. Gokhale, A. Javadi-Abhari, N. Earnest, Y. Shi, and F. T. Chong, “Op-
computation’s CQV. The QVA offers a nuanced, detailed timized quantum compilation for near-term algorithms with openpulse,”
method to estimate the failure rate of a given compiled cir- in Proc. 53rd Annu. IEEE/ACM Int. Symp. Microarchitecture, 2020,
pp. 186–200, doi: 10.1109/MICRO50266.2020.00027.
cuit, considering the effects of individual gate errors, their [15] I. Q. Group, “Open-source quantum development,” Accessed: Apr.
cumulative influence, and the unique properties of quantum 16, 2021. [Online]. Available: https://qiskit.org/
gates such as CNOT. To establish the efficacy of QVA, we [16] L. K. Grover, “A fast quantum mechanical algorithm for database search,”
subjected it to rigorous validation on cutting-edge quantum in Proc. 28th Annu. ACM Symp. Theory Comput., 1996, pp. 212–219,
doi: 10.1145/237814.237866.
machines using well-known benchmarks. The results demon- [17] IBM, “IBM quantum,” Accessed: Apr. 16, 2021. [Online]. Available:
strated that QVA consistently outperformed the prevalent SR https://quantum-computing.ibm.com/
estimator, ESP, and showcased linear scalability. On average, [18] K. Kechedzhi et al., “Effective quantum volume, fidelity and
computational cost of noisy quantum processing experiments,”
our model exhibited a sixfold reduction in the relative pre- Future Gen. Comput. Syst., vol. 153, pp. 431–441, Apr. 2023, doi:
diction error rate compared to ESP. Such accuracy not only 10.1016/j.future.2023.12.002.
bolsters confidence in our method but also has far-reaching [19] R. LaRose et al., “Mitiq: A software package for error mitigation
implications at both the hardware and software levels. on noisy quantum computers,” Quantum, vol. 6, 2022, Art. no. 774,
doi: 10.22331/q-2022-08-11-774.
[20] T. LeCompte, F. Qi, and L. Peng, “Robust cache-aware quantum processor
layout,” in Proc. Int. Symp. Reliable Distrib. Syst., 2020, pp. 276–287,
ACKNOWLEDGMENT doi: 10.1109/SRDS51746.2020.00035.
The views and conclusions contained in this document are [21] T. LeCompte, F. Qi, X. Yuan, N.-F. Tzeng, M. H. Najaf, and L. Peng,
those of the authors and should not be interpreted as repre- “Graph neural network assisted quantum compilation for qubit allo-
cation,” in Proc. ACM Great Lakes Symp. VLSI, 2023, pp. 415–419,
senting the official policies, either expressed or implied, of doi: 10.1145/3583781.3590300.
the U.S. Government. The U.S. Government is authorized to [22] T. LeCompte, F. Qi, X. Yuan, N.-F. Tzeng, M. H. Najafi, and L. Peng,
reproduce and distribute reprints for Government purposes “Machine learning-based qubit allocation for error reduction in quantum
notwithstanding any copyright notation herein. circuits,” IEEE Trans. Quantum Eng., vol. 4, 2023, Art. no. 3101414,
doi: 10.1109/TQE.2023.3301899.
[23] G. Li, Y. Ding, and Y. Xie, “Tackling the qubit mapping problem
for NISQ-era quantum devices,” in Proc. 24th Int. Conf. Architec-
REFERENCES tural Support Program. Lang. Operating Syst., 2019, pp. 1001–1014,
[1] “IBM quantum systems,” Accessed: Jul. 28, 2022. [Online]. Available: doi: 10.1145/3297858.3304023.
https://quantum-computing.ibm.com/services?systems=all [24] G. Li, A. Wu, Y. Shi, A. Javadi-Abhari, Y. Ding, and Y. Xie, “Pauli-
[2] N. Acharya and S. M. Saeed, “A lightweight approach to detect ma- hedral: A generalized block-wise compiler optimization framework for
licious/unexpected changes in the error rates of NISQ computers,” quantum simulation kernels,” in Proc. 27th ACM Int. Conf. Archi-
in Proc. IEEE ACM Int. Conf. Comput. Aided Des., 2020, pp. 1–9, tectural Support Program. Lang. Operating Syst., 2022, pp. 554–569,
doi: 10.1145/3400302.3415684. doi: 10.1145/3503222.3507715.
[3] N. Acharya and S.-M. Saeed, “Automated flag qubit insertion for reliable [25] J. Liu and H. Zhou, “Reliability modeling of NISQ-era quantum com-
quantum circuit output,” in Proc. IEEE Comput. Soc. Annu. Symp., 2021, puters,” in Proc. IEEE Int. Symp. Workload Characterization, 2020,
pp. 431–436, doi: 10.1109/ISVLSI51109.2021.00085. pp. 94–105, doi: 10.1109/IISWC50251.2020.00018.
[4] F. Arute et al., “Quantum supremacy using a programmable super- [26] E. Magesan, J. Gambetta, and J. Emerson, “Robust randomized
conducting processor,” Nature, vol. 574, no. 7779, pp. 505–510, 2019, benchmarking of quantum processes,” 2010, arXiv:1009.3639, doi:
doi: 10.1038/s41586-019-1666-5. 10.1103/PhysRevLett.106.180504.
[5] J. Clarke and F. K. Wilhelm, “Superconducting quantum bits,” Nature, [27] E. Magesan, J. M. Gambetta, and J. Emerson, “Characterizing quantum
vol. 453, no. 7198, pp. 1031–1042, 2008, doi: 10.1038/nature07128. gates via randomized benchmarking,” Phys. Rev. A, vol. 85, no. 4, 2012,
[6] L. DiCarlo et al., “Demonstration of two-qubit algorithms with a super- Art. no. 042311, doi: 10.1103/PhysRevA.85.042311.
conducting quantum processor,” Nature, vol. 460, no. 7252, pp. 240–244, [28] J. Majer et al., “Coupling superconducting qubits via a cavity bus,” Nature,
2009, doi: 10.1038/nature08121. vol. 449, no. 7161, pp. 443–447, 2007, doi: 10.1038/nature06184.

3100411 VOLUME 5, 2024

 IEEE Transactions on
Qi et al.: QUANTUM VULNERABILITY ANALYSIS TO GUIDE ROBUST QUANTUM COMPUTING SYSTEM DESIGN uantum Engineering

[29] A. Matsuo, W. Hattori, and S. Yamashita, “Reducing the overhead of [39] J. Poyatos, J. I. Cirac, and P. Zoller, “Complete characterization of a
mapping quantum circuits to IBM Q system,” in Proc. IEEE Int. Symp. quantum process: The two-bit quantum gate,” Phys. Rev. Lett., vol. 78,
Circuits Syst., 2019, pp. 1–5, doi: 10.1109/ISCAS.2019.8702439. no. 2, 1997, Art. no. 390, doi: 10.1103/PhysRevLett.78.390.
[30] N. Moll et al., “Quantum optimization using variational algorithms on [40] J. Preskill, “Quantum computing in the NISQ era and beyond,” Quantum,
near-term quantum devices,” Quantum Sci. Technol., vol. 3, no. 3, 2018, vol. 2, 2018, Art. no. 79, doi: 10.22331/q-2018-08-06-79.
Art. no. 030503, doi: 10.1088/2058-9565/aab822. [41] S. Resch, S. Tannu, U. R. Karpuzcu, and M. Qureshi, “A day in the life of
[31] P. Murali, J. M. Baker, A. Javadi-Abhari, F. T. Chong, and M. Martonosi, a quantum error,” IEEE Comput. Archit. Lett., vol. 20, no. 1, pp. 13–16,
“Noise-adaptive compiler mappings for noisy intermediate-scale quantum Jan.–Jun. 2020, doi: 10.1109/LCA.2020.3045628.
computers,” in Proc. 24th Int. Conf. Architectural Support Program. Lang. [42] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini,
Operating Syst., 2019, pp. 1015–1029, doi: 10.1145/3297858.3304075. “The graph neural network model,” IEEE Trans. Neural Netw, vol. 20,
[32] P. Murali, N. M. Linke, M. Martonosi, A. J. Abhari, N. H. Nguyen, no. 1, pp. 61–80, Jan. 2008, doi: 10.1109/TNN.2008.2005605.
and C. H. Alderete, “Architecting noisy intermediate-scale quantum com- [43] Y. Shi et al., “Optimized compilation of aggregated instructions
puters: A real-system study,” IEEE Micro, vol. 40, no. 3, pp. 73–80, for realistic quantum computers,” in Proc. 24th Int. Conf. Architec-
May/Jun. 2020, doi: 10.1109/MM.2020.2985683. tural Support Program. Lang. Operating Syst., 2019, pp. 1031–1044,
[33] P. Murali, N. M. Linke, M. Martonosi, A.-J. Abhari, N. H. Nguyen, doi: 10.1145/3297858.3304018.
and C. H. Alderete, “Full-stack, real-system quantum computer [44] P. W. Shor, “Polynomial-time algorithms for prime factorization and dis-
studies: Architectural comparisons and design insights,” in Proc. crete logarithms on a quantum computer,” SIAM Rev., vol. 41, no. 2,
ACM/IEEE 46th Annu. Int. Symp. Comput. Archit., 2019, pp. 527–540, pp. 303–332, 1999, doi: 10.1137/S0036144598347011.
doi: 10.1145/3307650.3322273. [45] R. Stassi, M. Cirio, and F. Nori, “Scalable quantum computer with super-
[34] P. Murali, D. C. McKay, M. Martonosi, and A. Javadi-Abhari, “Software conducting circuits in the ultrastrong coupling regime,” npj Quantum Inf.,
mitigation of crosstalk on noisy intermediate-scale quantum computers,” vol. 6, no. 1, pp. 1–6, 2020, doi: 10.1038/s41534-020-00294-x.
in Proc. 25th Int. Conf. Architectural Support Program. Lang. Operating [46] S. Tannu and M. Qureshi, “Not all qubits are created equal: A case for
Syst., 2020, pp. 1001–1016, doi: 10.1145/3373376.3378477. variability-aware policies for NISQ-era quantum computers,” in Proc.
[35] M. A. Nielsen and I. Chuang, Quantum Computation and Quan- 24th Int. Conf. Architectural Support Program. Lang. Operating Syst.,
tum Information. MA, USA: Cambridge, Univ. Press, 2002, doi: 2019, pp. 987–999, doi: 10.1145/3297858.3304007.
10.1017/CBO9780511976667. [47] S. S. Tannu and M. Qureshi, “Ensemble of diverse mappings: Improving
[36] S. Nishio, Y. Pan, T. Satoh, H. Amano, and R. Van Meter, “Extracting reliability of quantum computers by orchestrating dissimilar mistakes,”
success from IBM’s 20-qubit machines using error-aware compilation,” in Proc. 52nd Annu. IEEE/ACM Int. Symp. Microarchitecture, 2019,
ACM J. Emerg. Technol. Comput. Syst., vol. 16, no. 3, pp. 1–25, 2020, pp. 253–265, doi: 10.1145/3352460.3358257.
doi: 10.1145/3386162. [48] S. S. Tannu and M. K. Qureshi, “Mitigating measurement errors in
[37] D. Oliveira, E. Giusto, B. Baheri, Q. Guan, B. Montrucchio, and P. Rech, quantum computers by exploiting state-dependent bias,” in Proc. 52nd
“A systematic methodology to compute the quantum vulnerability factors Annu. IEEE/ACM Int. Symp. Microarchitecture, 2019, pp. 279–290,
for quantum circuits,” IEEE Trans. Dependable Secure Comput., early doi: 10.1145/3352460.3358265.
access, Sep. 12, 2023, doi: 10.1109/TDSC.2023.3313934.
[38] T. Patel, B. Li, R. B. Roy, and D. Tiwari, “UREQA: Leveraging operation-
aware error rates for effective quantum circuit mapping on NISQ-era quan-
tum computers,” in Proc. USENIX Annu. Tech. Conf., 2020, pp. 705–711.
[Online]. Available: https://www.usenix.org/system/files/atc20-patel.pdf

VOLUME 5, 2024 3100411