www.nature.

com/scientificreports

OPEN An optimizing method

for performance and resource
utilization in quantum machine
learning circuits
Tahereh Salehi1, Mariam Zomorodi1,2*, Pawel Plawiak2,3, Mina Abbaszade4 & Vahid Salari5,6
Quantum computing is a new and advanced topic that refers to calculations based on the principles
of quantum mechanics. It makes certain kinds of problems be solved easier compared to classical
computers. This advantage of quantum computing can be used to implement many existing problems
in different fields incredibly effectively. One important field that quantum computing has shown great
results in machine learning. Until now, many different quantum algorithms have been presented to
perform different machine learning approaches. In some special cases, the execution time of these
quantum algorithms will be reduced exponentially compared to the classical ones. But at the same
time, with increasing data volume and computation time, taking care of systems to prevent unwanted
interactions with the environment can be a daunting task and since these algorithms work on machine
learning problems, which usually includes big data, their implementation is very costly in terms of
quantum resources. Here, in this paper, we have proposed an approach to reduce the cost of quantum
circuits and to optimize quantum machine learning circuits in particular. To reduce the number of
resources used, in this paper an approach including different optimization algorithms is considered.
Our approach is used to optimize quantum machine learning algorithms for big data. In this case, the
optimized circuits run quantum machine learning algorithms in less time than the original ones and by
preserving the original functionality. Our approach improves the number of quantum gates by 10.7%
and 14.9% in different circuits respectively. This is the amount of reduction for one iteration of a given
sub-circuit U in the main circuit. For cases where this sub-circuit is repeated more times in the main
circuit, the optimization rate is increased. Therefore, by applying the proposed method to circuits with
big data, both cost and performance are improved.

In recent years, the phenomenon of quantum computing has received global ­attention1. Quantum computational
theory goes back to the works by Feynman and Deutsch in the ­1980s2 and after that many new quantum comput-
ing algorithms have been proposed. Machine learning is the science and art of building computers to learn from
data how to solve problems instead of explicitly programming. So, Machine learning and quantum computing
are two very important research areas, and by combining these two areas, new solutions for today’s challenges
are ­proposed3. There are some challenges for implementing quantum machine learning algorithms due to the
processing of large datasets. One usual way to meet these challenges is to arrange these algorithms in the cloud
system. With the help of the computing power of the cloud system, the problems are partially solved. However,
data storage and management in heterogeneous distributed networks had a number of other problems. Physicists
take a different method of quantum computing by exploiting the “Superposition” and “Entanglement” p ­ roperties9.
This solution has particularly increased the speed of solving certain problems compared to classical ­algorithms4,5.
Problem-solving is very different in quantum and classical systems. In fact, some problems that can be solved in the
­ ours6. On the other
classical system in several years, it is known that can be solved in a quantum system in a few h

1
Department of Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran. 2Department of
Computer Science, Faculty of Computer Science and Telecommunications, Cracow University of Technology,
Krakow, Poland. 3Institute of Theoretical and Applied Informatics, Polish Academy of Sciences, Gliwice,
Poland. 4Department of Physics, Isfahan University of Technology, Isfahan 84156‑83111, Iran. 5Institute for
Quantum Science and Technology, and Department of Physics and Astronomy, University of Calgary, Calgary,
T2N 1N4, Alberta, Canada. 6Basque Center for Applied Mathematics (BCAM), Bilbao, Spain. *email: zomorodi@
pk.edu.pl

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 1

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 1.  (a) An example for a quantum circuit: each line represents a qubit, and operations (gates) are applied
on them. (b) Representation of unitary two-qubit gates in the quantum circuit with matrix f­ ormalism17.

hand, in recent years many types of research have been carried out on the subject of big data. The challenge is the
inefficiency of the computations of classical machine learning algorithms and metaheuristics for processing such
a large volume of d ­ ata7–9. The unit of quantum processing is the “quantum bit” or “qubit”. One of the capabilities
of a quantum computer is that by increasing the number of qubits of a quantum computer, the processing power
improves ­exponentially10. Quantum algorithms usually express computations by primitive quantum gates. There
are different approaches to implementing these algorithms. Therefore, it is useful to find an implementation using
the least resource numbers, especially for large-scale quantum circuits with complex designs. To this end, we apply
optimization methods which is a fundamental task in almost all areas of quantum computing science, including
monolithic and distributed quantum circuits 10–13. This work has developed and implemented a framework for
quantum circuit optimization algorithms to optimize the desired circuits which are designed particularly for
machine learning tasks. We also show how to optimize the repetition of quantum circuits and reduce the required
resources for large-scale quantum circuits. While the original functionality of the algorithm is preserved, the final
quantum circuit has fewer time steps, execution time, and quantum cost compared to the original circuit. As input,
we assumed that the quantum circuit (QC) consists of a set of quantum gates with a certain number of 2 qubits .
The ultimate goal of optimizing the quantum circuit of a machine learning algorithm is to reduce the number of
gates, time steps, and quantum cost. The quantum cost of a circuit is the number of 1×1 and 2×2 quantum gates
­ esign14. For this purpose, this paper proposes a method to optimize the quantum cost of machine learning
in its d
algorithms. In principle, it can be said that the operations involved in quantum machine learning circuits can be
large and so it is worth reducing them. Quantum circuits typically use single-qubit and two-qubit gates such as
NOT, Hadamard, and rotation, and also two-qubit CNOT gates. If there are three-qubit gates such as Bridge, and
Swap, and multi-qubit gates, we decomposed them into single-qubit and two-qubit gates in a preprocessing step.
In “Quantum gates and circuits”, we discussed related work in the field of the quantum computation systems for
the machine learning algorithm, as well as optimization algorithms for the quantum circuits. Then, in “Quantum
memory”, the proposed method is explained and at the end, in “Related work” and “Quantum circuits optimiza-
tion techniques”, our results and discussion are presented and we conclude the paper.

Quantum gates and circuits

Quantum circuit is made up of a combination of Von Neumann and classical a­ rchitecture15, which is executed in
the quantum processor. Quantum circuits are shown in such a way that the desired gates are on vertical lines, each
of which represents a basic operation. The running time is calculated in these circuits from left to ­right16 using time
steps of the quantum circuit. Quantum gates can be represented as 2n dimensional matrices that contain the ampli-
tude of the fundamental states 2n of an n-dimensional quantum system. Figure 1 shows some examples of quantum
circuits and their corresponding gates in matrix representation. Figure 1a shows a sample quantum circuit, and
­ atrices17. Time advances from left to right.
Fig. 1b represents the unitary two-qubit gates with their corresponding m

Quantum memory
The memory of a classic computer can be easily built by writing an arbitrary bit string in any position. Classi-
cal memory performance is not optimal for processing big data. In order to solve the problem of normal and
associative memory capacity, quantum memory has been used successfully. In many applications of quantum
computers, a quantum register is used instead of classical memory to simulate a physical system. This quantum
memory consists of a qubit state tensor in multidimensional Hilbert space that is first prepared in a simple state.
For a quantum memory consisting of two qubits |q0 � and |q1 �, its state |qR � is given by Eq. (1). The symbol ⊗ is
the tensor operation. For example, for a 4-qubit quantum register, its state is represented as Eq. (2), where the
probability of measuring each of its base states is as different four terms illustrated in Eq. (3)18,19. Giovannetti
et al.20 have demonstrated that how classical data is presented in the language of quantum mechanics through
the study of quantum random access memory (QRAM).

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 2

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

|qR � = |q0 � ⊗ |q1 � = |q0 �|q1 � = |q0 q1 � (1)

|qR � = c0 |00� + c1 |01� + c2 |10� + c3 |11� (2)

|c0 |2 + |c1 |2 + |c2 |2 + |c3 |2 = 1 (3)

Implementing machine learning algorithms in quantum computing has two advantages: storage scale and
high execution speed of ­algorithm9. By exploiting the superposition property, it is possible that quantum storage
be reduced exponentially. According to Eqs. (4) and (5), all binary numbers from the set {0, 1, . . . , i, . . . , 2n − 1}
are placed in n qubit quantum memory with probability |ci |2 for state |i�.
n −1
2
|φ1 φ2 · · · φn � = ci |i� (4)
i=0

where
n −1
2
|ci |2 = 1 (5)
i=0

In the classical system for Eq. (6), the operation must be repeated 2n times. But in the quantum system, the
system can examine all computational states for the variables simultaneously, assuming that the Uf operator
understands the function f(x)9.
2n −1 2n −1 2n −1
 
1  1  1 
Uf √ |x� = √ Uf |x� = √ |f (x)� (6)
2n x=0 2n x=0 2n x=0

Related work
In this section, we first present some works done in implementing a quantum circuit for quantum machine
learning algorithms and then methods for quantum circuit optimization are presented.

Quantum circuits of machine learning algorithms. Recently, quantum machine learning is consid-
ered as a suitable solution to increase the speed of execution of algorithms. This method has led to the introduc-
tion of various quantum algorithms for machine learning using quantum features. In this paper, we first examine
the K nearest neighbor algorithm. In this regard, Lioyd et al.21 and Wiebe et al.22 use similar approaches such
as quantum amplitude estimation or Grover a­ lgorithm23 to obtain the quantum state of the nearest neighbor
algorithm. In the next method for implementing the nearest neighbor algorithm, Buhrman et al.24 use quantum
parallelism and the test circuit to calculate the distance between two vectors and provide a quantum solution.
The Euclidean distance can be calculated as Euclidean distance= ((2 − 2|�x|y�|)) . The next method in K near-
est neighbor by Ruan et al.25 is used in document classification, image classification, etc. It works based on the
size of the Hamming distance. A natural vector is defined as a bit vector with a hash function and then converted
to an equivalent quantum state, after that, then the input vector bits are compared with the training vector. The
number of different properties is counted by the Kaye ­circuit25 and the distance between the two vectors is esti-
mated. The next algorithm, the support vector machine (SVM), is a supervised algorithm developed by Arodz
and ­Saeedi26 and also by Rebentrost et al.27, which classifies vectors in a specific space based on training data.
In comparison with the classical support vector machine for binary classification, they achieved a logarithmic
acceleration. These methods use Grover algorithm and adiabatic algorithm. The next algorithm is the neural net-
work algorithm. Transfer learning is an interesting technique in neural networks in which a pre-trained model
is reused as an input model for a new task. One of the works in quantum neural networks presented in this field
was developed by Acar et al28 and uses the quantum transfer learning method. This method is a hybrid machine
learning method consisting of a classical network feature extractors and a diverse quantum classification circuit.
There are other works by Zen et al.29 that have used transmission learning toward scalable quantum neural
network states using transmission learning. A protocol was proposed ­in47 for machine translation based on
quantum long short term memory for translating a sentence from English to Persian. In another work, Mishra
et al.30 used the design and operation of a classical neural network and they designed a quantum neural network
capable of working on a 10 qubit system. By demonstrating network performance, they have tried to use the
basic principles of machine learning to manage data that can be used in cancer detection.

Quantum circuits optimization techniques

There are many different optimization methods for quantum circuits that aim to reduce the number of time steps,
gates, depth, and more. For this purpose, quantum circuits can be improved using different approaches in circuit
diagram models, or using circuit simplification rules. For example, the quantum circuit can be improved by the
quantum Karnaugh map ­method31, the exclusive sum of product ­method32 and etc. Using circuit simplification
rules, there are many different methods that improve circuits by specific rules. The following is a review of the
works in this area. Fan et al.33 proposed a quantum approximate optimization algorithm, which is a standard
method for combinatorial optimization with a gate-based quantum computer. The paper introduced a new Gibbs
objective function and demonstrated its superiority, and used the architecture of an Ansatz search algorithm

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 3

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

to search for the discrete space of a quantum circuit. These changes led to the improvement of various circuits.
Using this method, the median has been increased to 244.7% and 44.4% for the grid and complete graph models
of quantum computation. Median reduction in the number of two-qubit gates is 33.3% and 20.8%, respectively. In
another paper, Alam et al.34 proposed a method to accelerate the implementation of the quantum approximation
optimization algorithm (QAOA). First, a connection is made between the classical optimizer and the quantum
computer, and then two parameters named δ and β , with initial values of zero, are inserted into the loop. The
classical optimizer for randomly defined variables initially set to some random values. If the values are not ideal,
it establishes a connection to the quantum computer. increases the depth of the circuit, which is not good and
should be reduced. For this reason, to determine the appropriate distance for the parameters, artificial intelligence
techniques are used to achieve the desired result with the acceleration in the process. This method shows that
the number of optimization iterations can be reduced 44.9% on average for 264 graphs. Haner et al.15 optimized
the circuit using the Hoare ­triples35. This method checks the accuracy of the execution of specific programs.
For each circuit level, a pre-condition defines conditions and post-conditions, and based on the previous level
condition, the authors can decide on the operating conditions for the next level operation. When using a Hoare-
based optimization strategy, the circuit depth decreases for n ≥ 2, according to relation (4(n − 2) + n)/n. In the
next method of Childs and ­Maslov16, the automatic optimization of large circuits is accomplished using iterative
parameters. This method also preserves the main structure of the algorithm and performs better optimizations
than state-of-the-art approaches. In fact, it uses a set of exploratory laws that reduces the number of gates. This
technique first displays the quantum circuit as a netlist and then preprocesses and simplifies the circuit. Then, it
divides the circuit into sub-circuits and optimizes the sub-circuits according to the rules 1-4. I­ n36, Abdessaied
et al. used several algorithms to synthesize reversible functions to quantum circuits and to reduce the number of
Hadamard gates. This reduction of the Hadamard gate, reduces the number and depth of T gates, which improves
the combined gates. By applying this method, the authors improved the number and depth of T gate by 88% more
than other optimization methods. One other approach for quantum circuit optimization is based on ZX-calculus
which is a graphical language for expressing quantum ­computation49. The optimization approach uses the rules
of the ZX-calculus for simplifying ZX- ­diagrams50. The authors show that their simplification procedure works
well when there are few non-clifford gates in the original circuit. Using different quantum circuit optimization
techniques, the aim of this paper is to improve the performance of quantum machine learning circuits and to
reduce their cost. To this end we optimized the quantum machine learning circuits in terms of quantum gates
and time steps.

Methods
Implementing machine learning algorithms with big data in quantum systems is a major challenge due to the
excessive increase in the number of gates, the depth of the circuit, and the execution time of the algorithm.
Optimizing quantum circuits is an effective way to overcome these problems. In this section, the details of
the optimization algorithm for quantum machine learning circuits are explained. This method is then used to
optimize the quantum circuits of two machine learning algorithms, transmission learning and neural networks.
Initially, in the preprocessing step, the quantum circuit represented as a list of gates that are applied sequentially.
The following transformation rules are then applied to optimize the quantum machine learning circuits.
Rule 1: First, if there is a NOT gate in the circuit, the next gate is checked. In this case, there are three differ-
ent possibilities for the next ­gates16:

• If the next gate is a TOFFOLI gate: in this case the control qubit of the TOFFOLI is reversed and the NOT
gate is removed.
• If the next gate is a NOT gate: in this case the two NOT gates are removed.
• If the next gate is a CNOT gate: in this case the control qubit is reversed and the NOT gate is removed.

Rule 2: Remove gates that are directly adjacent to their inverse. In a two-qubit gate, it is usually possible to
simplify or eliminate the gate in the form of quantum circuits by moving it between the gates. In fact, for each U
gate in the circuit, the optimizer searches for an instance of U †. If present, U is successfully canceled with some
instances of U †.
Rule 3: For two rotation gates RZ (θi ) and RZ (θj ) that have a shared control line, According to Eq. (7), we can
merge two ­rotations37. For example, in Fig. 2 two rotation gates RZ (θ1 ) and RZ (θ4 ) can be c­ ombined16:
R(θ1 ) · R(θ2 ) = R(θ1 + θ2 ) (7)

R(θ1 ) · R(θ2 ) = R(θ2 ) · R(θ1 ) (8)

Rule 4: Because many quantum algorithms can be described using Swap and Bridge ­gates38, we transform
them into the equivalent circuits consisting of CNOT gates on two consecutive qubits. Figure 3a is the equivalent
circuit of Swap gate and Fig. 3b is the equivalent circuit of Bridge gate. By breaking down multi-qubit gates into
simpler gates, the resulting circuit performs better when using other r­ ules38.
In some cases, the gates that can be merged or removed are not placed side by side. By moving the gates
according to Figs. 4 and 5, the gates are placed side by side and so they can be merged and removed by the above
rules. . For this purpose, the Sympy ­library39 is used. Sympy is an open source Python library for symbolic
mathematics. With the help of this library, complex quantum circuits can be transformed into simple ones. One
of the interesting features of this library is producing equivalent circuits. In fact, using library different gates in

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 4

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 2.  Integration of two rotation g­ ates16.

Figure 3.  (a) Equivalent circuits of Swap gates. (b) Equivalent circuits of Bridge g­ ates38.

Figure 4.  A quantum circuit and its equivalent.

Figure 5.  (a) Commuting of the rotation gates and the CNOT gate. (b) Commuting of two CNOT gates in two
different ­circumstances38.

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 5

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 6.  The flowchart of the steps of the optimization approach using rules 1–4.

the circuit are moved around and all locations where the gates can be placed are examined. Then, the rules 1-4
are re-examined by the algorithm and the circuit is simplified if conditions pass.
These defined rules are applied in a loop until no further improvement is obtained. Algorithm 1 and Fig. 6
present the steps of our optimization approach using the above rules. Using this framework, we optimized the
quantum machine learning circuit of a classification task for medical diagnosis using quantum transfer ­learning28.
This circuit has been tested in several real quantum processors as well as various simulators. This quantum
circuit aims at distinguishing a sick person from a healthy person based on computed tomography images. The
circuit consists of four steps: The Hadamard gate is first applied to all qubits and then with the help of U operator
defined ­in28, the classical data is encoded and then entanglement is created. The dotted box of Fig. 7 shows one
application of this operator. Finally, the qubits are measured. Figure 7 demonstrates the quantum circuit of this
quantum machine learning algorithm with only one repetition of the sub-circuit.

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 6

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 7.  A quantum circuit with a quantum transfer learning m

­ ethod28.

Results and discussion

In this section, the experimental results for different quantum circuits and quantum machine learning circuits
are presented. The file of these circuits are mostly in QASM format, received as input by Python language and
then converted into a matrix by the Qiskit library of ­Python46. Then the proposed optimization techniques are
applied to them. The main criterion in evaluating the proposed method is the comparison of the number of gates
before and after optimization. Also, another criterion we considered in this work is the execution time of the
quantum circuits and the amount of time step reduction. Qiskit library was used in IBM Q quantum computers
for output verification. The optimization is then performed and the circuit resulting from the optimization was
given to the simulator and the number of gates is calculated after the simulation. Obviously, in order to obtain
the simulation results, the pre-optimization and post-optimization simulations must be the same. As the number
of gates decreases from the initial value after applying the optimization model, the speed of the algorithm and

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 7

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Pre-optimization After Optimization Improvement(%)

Execusion Execution Execution
Quantum time Imp. #of Time time Imp. #of Time time Imp. #of Time
circuit #of qubits (S) steps #of Gates (S) steps #of Gates (S) steps #of Gates Speed up References
VQC 3 6.4 13 21 6.3 13 18 1.56 0 14.28 0.1 40

Q transfer 29
4 6.8 13 28 6.2 13 25 8.82 – 10.71 0.6
learning
Q neural 28
10 6.9 17 67 6.6 16 57 4.34 5.88 14.92f 0.3
network
41
Grover 1 4 6.4 26 41 6.2 16 21 3.12 38.46 48.78 0.2
41
Grover 2 4 6.7 21 39 5.9 17 27 11.94 19.09 30.76 0.8
42
QAOA 2 6.7 7 11 6.2 5 7 7.46 28.57 36.36 0.5
test circuit K 43
3 6.5 14 26 6.2 14 24 4.6 0 7.7 0.3
Means
44
TNN 4 6 10 27 5.9 8 14 1.7 20 48.14 0.1
45
KNN 4 7.8 20 33 4.7 8 13 39.74 60 60.60 3.1

Table 1.  Results of implementation of the proposed method on different quantum circuits.

its implementation cost improve. In this case, the proposed model will be a more efficient model. In order to
verify our approach, we first tested our approach on different general quantum circuits and the results are shown
in Table 1. In this table, each column is the corresponding quantum circuit and for each circuit we showed the
improvement caused by our optimization approach.
Also, in Table 2 the comparison between our proposed approach and other works in the literature ZX-
calculus50, ­AQCEL51, ­tket52, and Q­ uilc53 are presented. It can be seen from this table that our approach works
better in terms of circuit depth compared to other approaches and also in many circuits it is better in terms of
the number of 2-qubit gates, while the execution time of our method is better than all other approaches.
In the proposed method, assuming that the number of time steps is N and the number of qubits is Q, the
time complexity of the algorithm is obtained as O(NQ). As shown in Tables 1 and 2 applying our method to a
variety of quantum circuits reduces the number of gates, time steps, and execution time of the quantum circuits
significantly. At the second part of the experiments, our optimization approach was applied on the quantum
machine learning circuits. One of these circuits uses transmission learning method for a potential application
in medical diagnosis. By applying the proposed method to the above quantum circuit, only the U-shaped part of
the circuit improves as shown in Fig. 8. In Fig. 8a it can be seen that the original circuit ­from28 has 28 quantum
gates. Figure 8b shows the improved circuit diagram with 10.7% reduction in the number of gates. This is the
amount of quantum cost reduction for one repetition below the U-circuit in the main circuit. For cases where
this sub-circuit is repeated many times in the main circuit, the rate of improvement increases. In this case, by
applying the proposed method on circuits with big data, desirable results will be obtained. The results of the
implementation of the proposed method on the quantum circuit of transfer learning are shown in Fig. 9 before
and after optimization.
We verified the outputs in IBM Q and the results are demonstrated in Fig. 10. Figure 10a is the output of the
original circuit and Fig. 10b is the output after we applied our optimization algorithm. Since the output is the
same in both cases, the transformation has done correctly.
The next quantum machine learning circuit that we used in this work is the quantum circuit of the neural
network for cancer d ­ etection30 which used the design and operation of a classical neural network but it is a
quantum neural network capable of working on a 10 qubit system. By demonstrating network performance, the
authors have tried to use the basic principles of machine learning to manage data. The graphical representation
of this circuit is shown in Fig. 11. Figure 11a shows the original circuit f­ rom30 which is implemented in 17 time
steps with 67 quantum gates. Figure 11b shows the improved circuit which in addition to a 14.9% reduction in
the number of gates, reduces its time steps to 16. The comparison result of applying the proposed method to
this circuit is shown in Fig. 12.
The output results of the circuits are shown in Fig. 13. Figure 13a is the output of the original circuit and
Fig. 13b is the output after we applied our optimization algorithm. Since the output is the same in both cases, it
shows that the transformation of the proposed optimization is correct.
The next quantum machine learning circuit that we used in this work is Fig. 14. The quantum repeater cir-
cuit is used as a test for the KNN algorithm i­ n45. The graphical representation of this circuit is shown in Fig. 14.
Figure 14a shows the original circuit, which is implemented in 20 time steps with 33 quantum gates. Figure 14b
shows the improved circuit, which in addition to a 60.60% reduction in the number of gates, reduces its time
steps to 8. The comparison result of applying the proposed method to this circuit is shown in Fig. 15.
We verified the outputs in IBM Q and the results are demonstrated in Fig. 16. Figure 16a is the output of the
original circuit and Fig. 16b is the output after we applied our optimization algorithm. Since the output is the
same in both cases, the transformation has done correctly.

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 8

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

Circuit name Method 2-qubit gate count Depth Time

ZX Calculus 5 5 0.08
Quilc 5 5 1.03
Graycode6-47 tket 5 5 0.06
AQCEL 5 5 2.08
Proposed method 5 5 0.04
ZX Caculus 8 10 0.76
Quilc 8 7 0.83
4gt11-84 tket 12 11 0.01
AQCEL 8 7 0.98
Proposed method 8 7 0.06
ZX Calculus 15 28 1.53
Quilc 18 17 1.64
4mod-v1-24 tket 26 25 0.05
AQCEL 19 22 0.87
Proposed method 28 16 0.03
ZX Calculus 42 48 1.43
Quilc 46 36 2.16
Decod24-bdd-294 tket 59 52 0.04
AQCEL 45 46 2.87
Proposed method 54 38 0.02
ZX Calculus 112 110 4.37
Quilc 135 78 16.32
Mini-alu-305 tket 178 112 0.09
AQCEL 172 95 0.32
Proposed method 168 90 0.05
ZX Calculus 23386 18380 215.06
Quilc 3632 2822 139.35
PF4_30_after
tket 3614 2798 5.33
Proposed method 3638 2798 4.86
ZX Calculus 3426 3620 145.15
Quilc 4130 3517 113.25
Dc2_222
tket 4130 3517 2.83
Proposed method 4131 3517 1.83
ZX Calculus 2295 2052 90.47
Quilc 2560 2102 2.64
Square_root_7
tket 3089 2102 2.56
Proposed mthod 2698 2098 1.43
ZX Calculus 3044 2168 56.08
Quilc 1182 1182 76.33
Pf4_20_before
tket 1182 1182 1.52
Proposed Method 1196 1194 1.39

Table 2.  Comparison results of the proposed method with state-of-the-art optimization methods on different
quantum circuits.

Conclusion
Realizing machine learning algorithms in a quantum system for big data is a real challenge but with remarkable
advantages of using quantum computers. In quantum circuits, as the number of gates increases, the number of
time steps and execution time is also increased, which is why optimizing quantum circuits is an effective way to
overcome these problems. In this study, a new general framework of quantum circuit optimization was presented
and in particular, quantum machine learning algorithms for big data were investigated in order to improve their
quantum circuit model which in turn leads to the improvement and reduction in the number of required quan-
tum computation resources. In fact, by applying the proposed method, quantum circuits were implemented in
less time than the original circuits, with the same functionality of the original design. In addition, applying this
method also reduces the quantum costs. Several quantum circuits with different functionality and algorithms
were used to evaluate the proposed method. The results of the improved circuits showed that the number of
quantum gate, the time steps, and the execution time in the evaluated circuits were reduced. In particular,
the proposed method was investigated on the quantum circuits of transfer learning and neural network. Our

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 9

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 8.  Demonstration of original and improved quantum transfer learning circuits. Diagram (a) shows the
non-optimal circuit and diagram (b) shows the improved circuit with reducing the number of gates by 10.7%.

Figure 9.  Demonstration of optimized and non-optimized diagrams of quantum transfer learning circuits.

Figure 10.  The simulation result of the output of the quantum transfer learning circuit before and after
optimization in (a) and (b), respectively. The results are identical.

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 10

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 11.  Demonstration of o­ riginal30, and (b) improved quantum neural network circuits used for cancer
diagnosis. The non-optimal circuit (a) is executed in 17 time steps, but the improved circuit, which has a
reduction of 14.9% in the number of gates, its time step is reduced to 16.

Figure 12.  Comparison between optimized and non-optimized quantum neural networks circuits.

approach reduced the number of the gates by 10.71% respectively in transfer learning circuit and also reduced the
number of time steps and the gate by 27.2% and 14.9% respectively in neural network circuit. More importantly,
this was the amount of reduction for one iteration of the U-subcircuit in the main circuit of the transfer learning
algorithm. So, for the cases where this sub-circuit was repeated more often in the main circuit, the optimization
is even more. So, by applying the proposed method on circuits with big data, better results would be obtained.

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 11

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 13.  The simulation result of the output of the quantum neural network circuit before and after
optimization in (a) and (b) respectively. The results are identical.

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 12

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 14.  (a) Demonstration of ­original30, and (b) improved quantum circuits test for the KNN algorithm.
The non-optimal circuit (a) is executed in 20 time steps, but the improved circuit, which has a reduction of
60.60% in the number of gates, its time step is reduced to 8.

Figure 15.  Comparison between optimized and non-optimized quantum circuits test for the KNN algorithm.

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 13

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 16.  (a) The simulation result of the output of the quantum circuits test for the KNN algorithm before
and after optimization in (a) and (b), respectively. The results are identical.

Data availibility
The datasets used and/or analysed during the current study available from the corresponding author on reason-
able request.

Received: 29 December 2021; Accepted: 13 September 2022

References
1. Humble, T. S., Thapliyal, H., Munoz-Coreas, E., Mohiyaddin, F. A. & Bennink, R. S. Quantum computing circuits and devices.
IEEE Des. Test. 36(3), 69–94 (2019).
2. Benenti, G., Casati, G., & Strini, G. Principles of Quantum Computation and Information-Volume II: Basic Tools and Special
Topics. World Scientific Publishing Company (2007).
3. Schuld, M. & Petruccione, F. Supervised learning with quantum computers (Springer, Berlin, 2018).
4. Grover, L.K. A fast quantum mechanical algorithm for database search. InProceedings of the twenty-eighth annual ACM sympo-
sium on Theory of computing pp. 212–219 (1996)
5. Humble, T. S., Thapliyal, H., Munoz-Coreas, E., Mohiyaddin, F. A. & Bennink, R. S. Quantum computing circuits and devices.
IEEE Des. Test. 36(3), 69–94 (2019).
6. Gyongyosi, L. & Imre, S. A survey on quantum computing technology. Comput. Sci. Rev. 1(31), 51–71 (2019).
7. Dang, Y., Jiang, N., Hu, H., Ji, Z. & Zhang, W. Image classification based on quantum K-Nearest-Neighbor algorithm. Quantum
Inf. Process. 17(9), 1–8 (2018).
8. Beheshti Roui, M., Zomorodi, M., Sarvelayati, M., Abdar, M., Noori, H., Pławiak, P., Tadeusiewicz, R., Zhou, X., Khosravi, A.,
Nahavandi, S. & Acharya, U. R. A novel approach based on genetic algorithm to speed up the discovery of classification rules on
GPUs. Knowl Based Syst. 231, 107419 https://​doi.​org/​10.​1016/j.​knosys.​2021.​107419 (2021).
9. Ruan, Y., Xue, X., Liu, H., Tan, J. & Li, X. Quantum algorithm for k-nearest neighbors classification based on the metric of ham-
ming distance. Int. J. Theor. Phys. 56(11), 3496–507 (2017).
10. Savchuk, M. M. & Fesenko, A. V. Quantum Computing: Survey and Analysis. Cybern. Syst. Anal. 55(1), 10–21 (2019).
11. Ghodsollahee, I., Davarzani, Z., Zomorodi, M., Pławiak, P., Houshmand, M. & Houshmand, M. Connectivity matrix model of
quantum circuits and its application to distributed quantum circuit optimization. Quantum Inf. Process. 20(7), 235https://​doi.​
org/​10.​1007/​s11128-​021-​03170-5(2021).
12. Gilyén, A., Arunachalam, S., & Wiebe, N. Optimizing quantum optimization algorithms via faster quantum gradient computa-
tion. In Proceedings of the Thirtieth Annual ACM-SIAM Symposium on Discrete Algorithms 2019 (pp. 1425–1444). Society for
Industrial and Applied Mathematics.
13. Daei, O., Navi, K. & Zomorodi-Moghadam, M. Optimized Quantum Circuit Partitioning.Int. J. Theor. Phys. 59(12), 3804-3820
https://​doi.​org/​10.​1007/​s10773-​020-​04633-8 (2020).
14. Thapliyal, H. & Ranganathan, N. Design of reversible sequential circuits optimizing quantum cost, delay, and garbage outputs.
ACM J. Emerg. Technol. Comput. Syst. (JETC). 6(4), 1–31 (2010).
15. Häner, T., Hoefler, T., & Troyer, M. Using Hoare logic for quantum circuit optimization. ArXiv e-prints. (2018).
16. Nam, Y., Ross, N.J., Su, Y., Childs, A.M., & Maslov, D. Automated optimization of large quantum circuits with continuous param-
eters. NPJ Quant. Inf. 4(1), 1–2 (2018).
17. Schuld, M., Sinayskiy, I. & Petruccione, F. An introduction to quantum machine learning. Contemp. Phys. 56(2), 172–85 (2015).
18. Hagouel, P.I., & Karafyllidis, I.G. Quantum computers: Registers, gates and algorithms. In 2012 28th International Conference on
Microelectronics Proceedings pp. 15-21. IEEE (2012).
19. Soklakov, A. N. & Schack, R. Efficient state preparation for a register of quantum bits. Phys. Rev. A 73(1), 012307 (2006).
20. Giovannetti, V., Lloyd, S. & Maccone, L. Quantum random access memory. Phys. Rev. Lett. 100(16), 160501 (2008).
21. Lloyd, S., Garnerone, S. & Zanardi, P. Quantum algorithms for topological and geometric analysis of data. Nat. Commun. 7(1),
1–7 (2016).
22. Wiebe, N., Granade, C., Ferrie, C. & Cory, D. Quantum Hamiltonian learning using imperfect quantum resources. Phys. Rev. A
89(4), 042314 (2014).
23. Grover, L.K. A fast quantum mechanical algorithm for database search. In Proceedings of the twenty-eighth annual ACM symposium
on Theory of computing pp. 212–219 (1996).
24. Buhrman, H., Cleve, R., Watrous, J. & De Wolf, R. Quantum fingerprinting. Phys. Rev. Lett. 87(16), 167902 (2001).
25. Kaye, P. Reversible addition circuit using one ancillary bit with application to quantum computing. arXiv preprint. arXiv:q
​ uant-p
​ h/​
04081​73 (2004).
26. Saeedi, S., & Arodz, T. Quantum sparse support vector machines. arXiv preprint arXiv:​1902.​01879 (2019).

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 14

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

27. Rebentrost, P., Mohseni, M. & Lloyd, S. Quantum support vector machine for big data classification. Phys. Rev. Lett. 113(13),
130503 (2014).
28. Acar, E. & Yilmaz, I. COVID-19 detection on IBM quantum computer with classical-quantum transfer learning. Turk. J. Electr.
Eng. Comput. Sci. 29(1), 46–61 (2021).
29. Zen, R. et al. Transfer learning for scalability of neural-network quantum states. Phys. Rev. E 101(5), 053301 (2020).
30. Mishra, N., Bisarya, A., Kumar, S., Behera, B.K., Mukhopadhyay, S., & Panigrahi PK. Cancer Detection Using Quantum Neural
Networks: A Demonstration on a Quantum Computer. arXiv preprint arXiv:​1911.​00504 (2019).
31. Bae, J. H., Alsing, P. M., Ahn, D. & Miller, W. A. Quantum circuit optimization using quantum Karnaugh map. Sci. Rep. 10(1), 1–8
(2020).
32. Basak, A., Sadhu, A., Das, K. & Sharma, K. K. Cost Optimization Technique for Quantum Circuits. Int. J. Theor. Phys. 58(9),
3158–79 (2019).
33. Li, L., Fan, M., Coram, M., Riley, P. & Leichenauer, S. Quantum optimization with a novel gibbs objective function and ansatz
architecture search. Phys. Rev. Res. 2(2), 023074 (2020).
34. Alam, M., Ash-Saki, A., & Ghosh, S. Accelerating quantum approximate optimization algorithm using machine learning. In 2020
Design, Automation & Test in Europe Conference & Exhibition (DATE) (pp. 686–689, 2020). IEEE.
35. Hoare, C. A. An axiomatic basis for computer programming. Commun. ACM 12(10), 576–80 (1969).
36. Abdessaied, N., Soeken, M., & Drechsler, R. Quantum circuit optimization by Hadamard gate reduction. In International Confer-
ence on Reversible Computation pp. 149–162. Springer, Cham (2014).
37. Zomorodi-Moghadam, M. & Navi, K. Rotation-based design and synthesis of quantum circuits. J. Circ. Syst. Comput. 25(12),
1650152 (2016).
38. Itoko, T., Raymond, R., Imamichi, T. & Matsuo, A. Optimization of quantum circuit mapping using gate transformation and com-
mutation. Integration. 1(70), 43–50 (2020).
39. Curry, M. Symbolic quantum circuit simplification in SymPy.
40. Variational Quantum Classifier - Syed Farhan (born-2learn.github.io)
41. Mandviwalla, A., Ohshiro, K., & Ji, B. Implementing Grover’s algorithm on the IBM quantum computers. In2018 IEEE International
Conference on Big Data (Big Data) (pp 2531–2537, 2018). IEEE.
42. Karalekas, P. J. et al. A quantum-classical cloud platform optimized for variational hybrid algorithms. Quant. Sci. Technol. 5(2),
024003 (2020).
43. Larose, R. Overview and comparison of gate level quantum software platforms. Quantum 25(3), 130 (2019).
44. https://​cds.​cern.​ch/​record/​27162​04/​plots.
45. LaBorde, M. L., Rogers, A. C. & Dowling, J. P. Finding broken gates in quantum circuits: Exploiting hybrid machine learning.
Quant. Inf. Process. 19(8), 1–8 (2020).
46. McKay, D.C., et al. Qiskit backend specifications for openqasm and openpulse experiments. arXiv preprint arXiv:​1809.​03452
(2018).
47. Abbaszade, M., Salari, V., Mousavi, S. S., Zomorodi, M. & Zhou, X. Application of quantum natural language processing for lan-
guage translation. IEEE Access. 30(9), 130434–48 (2021).
48. Salari, V., et. al. Quantum Face Recognition Protocol with Ghost Imaging. preprint: arXiv:​2110.​10088. [quant-ph]
49. Coecke, B. & Duncan, R. Interacting quantum observables: Categorical algebra and diagrammatics. New J. Phys. 13(4), 043016
(2011).
50. Duncan, R., Kissinger, A., Perdrix, S. & Van De Wetering, J. Graph-theoretic Simplification of Quantum Circuits with the ZX-
calculus. Quantum 4(4), 279 (2020).
51. Jang, W., Terashi, K., Saito, M., Bauer, C.W., Nachman, B., Iiyama, Y., Kishimoto, T., Okubo, R., Sawada, R., & Tanaka, J. Quantum
gate pattern recognition and circuit optimization for scientific applications. In EPJ Web of Conferences 2021 (Vol. 251, p. 03023).
EDP Sciences.
52. Sivarajah, S. et al. t| ket: a retargetable compiler for NISQ devices. Quant. Sci. Technol. 6(1), 014003 (2020).
53. Smith, R. S., Peterson, E. C., Skilbeck, M. G. & Davis, E. J. An open-source, industrial-strength optimizing compiler for quantum
programs. Quant. Sci. Technol. 5(4), 044001 (2020).

Acknowledgments
VS thanks the financial support from Natural Sciences and Engineering Research Council (NSERC) and the
National Research Council (NRC) of Canada.

Author contributions
T.S. contributed to implementation of the research and the main conceptual ideas, and performing the experi-
ments. M.Z. contributed to the original idea, took the lead in writing the manuscript with input from all authors,
and were in charge of overall direction. P.P. supervised and commented on the manuscript and contributed to
the final version. M.A. contributed to the final version of the manuscript and commented on the manuscript.
V.S. contributed to the analysis of the results and also contributed to the writing of the manuscript.

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to M.Z.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 15

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons licence, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.

© The Author(s) 2022

Scientific Reports | (2022) 12:16949 | https://doi.org/10.1038/s41598-022-20375-5 16

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

onlineservice@springernature.com