npj | computational materials Article

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences

[Link]

Quantum-inspired genetic algorithm for

designing planar multilayer photonic
structure
Check for updates
1 1 2 1 3
Zhihao Xu , Wenjie Shang , Seongmin Kim , Alexandria Bobbitt , Eungkyu Lee &
Tengfei Luo 1,4

Quantum algorithms are emerging tools in the design of functional materials due to their powerful
solution space search capability. How to balance the high price of quantum computing resources and
1234567890():,;
1234567890():,;

the growing computing needs has become an urgent problem to be solved. We propose a novel
optimization strategy based on an active learning scheme that combines the Quantum-inspired
Genetic Algorithm (QGA) with machine learning surrogate model regression. Using Random Forests
as the surrogate model circumvents the time-consuming physical modeling or experiments, thereby
improving the optimization efficiency. QGA, a genetic algorithm embedded with quantum mechanics,
combines the advantages of quantum computing and genetic algorithms, enabling faster and more
robust convergence to the optimum. Using the design of planar multilayer photonic structures for
transparent radiative cooling as a testbed, we show superiority of our algorithm over the classical
genetic algorithm (CGA). Additionally, we show the precision advantage of the Random Forest (RF)
model as a flexible surrogate model, which relaxes the constraints on the type of surrogate model that
can be used in other quantum computing optimization algorithms (e.g., quantum annealing needs
Ising model as a surrogate).

In the pursuit of continuous technological advancement in the field of network model would be trained as a surrogate to mimic computationally
materials science, the design and discovery of novel functional materials are expensive physical simulations for evaluating the topology-property rela-
always at the forefront of innovation. In practice, the performance of dif- tionship. Then, a generative model, such as Generative Adversarial Net-
ferent functional materials is dependent on many design factors, such as works (GANs)5,7,8 or Variational Autoencoders (VAEs)9–11, is used to
geometrical features, composition, processing conditions, and environ- inversely retrieve the design based on material properties12–14.
mental factors, which lead to large design spaces1–3. Traditional methods, Another widely used approach is to use machine learning models as
including experiments and simulations, are usually too time-consuming surrogate models and combine them with global optimization algorithms.
and expensive to comprehensively search the extremely large design spaces. These surrogate model-based algorithms can effectively address the one-to-
Therefore, for general functional material design problems (such as alloy many challenge in many inverse design problems, i.e., there are multiple
materials, optical materials, etc.), although comprehensive physical models non-unique solutions to the same design target12. Compared to direct neural
have been established, traditional design methods can still only explore a network inverse design, this approach heavily relies on the performance of
small part of the design space around the configurations that have been well the global optimization algorithm to iteratively search the optimization
investigated. space. Some classical optimization algorithms such as Genetic Algorithm
With the development of computational science and data science, (GA, or classical GA) and Simulated Annealing algorithm (SA) have been
machine learning algorithms begin to transform the field of material design. widely applied for material design and optimization15–19. These algorithms
Taking the design of optical structures as an example, a series of deep can effectively explore the design space with limited information20. With the
learning-based algorithms combining forward modeling and inverse design emergence and development of quantum computing, a variety of sophis-
have been proposed and applied in practice4–6, by which an artificial neural ticated, high-performance quantum algorithms have been employed to

1
Department of Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN, 46556, USA. 2National Center for Computational Sciences, Oak
Ridge National Laboratory, Oak Ridge, TN, 37830, USA. 3Department of Electronic Engineering, Kyung Hee University, Yongin-si, Gyeonggi-do, 17104, Republic of
Korea. 4Department of Chemical and Biomolecular Engineering, University of Notre Dame, Notre Dame, IN, 46556, USA. e-mail: eleest@[Link]; tluo@[Link]

npj Computational Materials | (2024)10:257 1

[Link] Article

tackle complex and large-scale optimization problems, facilitating the for solving binary combinatorial optimization problems with complex
material designing with surrogate-based algorithms. One example is the searching spaces.
Quantum Annealing (QA) algorithm, which provides highly accurate
solutions for a specific kind of combinatorial optimization known as Results
Quadratically Constrained Binary Optimization (QUBO)21–23. However, for The Quantum Genetic Algorithm
any problems that have complex optimization landscapes, utilizing QUBO In QGA, a quantum chromosome, composed of multiple qubits,
as the surrogate model and solving it with QA has been the necessary represents a potential solution. This algorithm mimics the evolution of
strategy since QA can only work with QUBO-like models (e.g., the Ising physical qubits undergoing quantum logic gates, where “qubits” and
model). The errors introduced by ignoring higher-order terms beyond the “quantum gates” refer to mathematical-quantum information and
second order in the surrogate model require many iterations between operators within the complex Hilbert space, respectively. The algo-
QUBO training/retraining, QA, and data collection before convergence is rithm manipulates these chromosomes using quantum gates, analo-
achieved24. gous to genetic operations of crossover and mutation. Leveraging
Motivated by quantum computing theories, some heuristic algorithms superposition, a quantum chromosome can represent multiple solu-
were proposed as the enhanced solution for classical algorithms. One of the tions simultaneously, vastly increasing the algorithm’s ability to
representative examples is the Quantum-inspired GA (QGA). The QGA explore a large search space. The use of quantum entanglement can
was first introduced by Narayanan et al. 25. It is distinguished by its smaller introduce correlations between qubits, enabling the algorithm to
population size, rapid convergence speed, robust global optimization cap- maintain a higher diversity of solutions and avoid premature con-
abilities, and strong resilience to variations in problem specifications26,27. vergence on local optima. These characteristics make QGA particularly
QGA integrates the principles of quantum computing with the robustness of suitable for complex optimization problems where traditional genetic
classical genetic algorithms, thereby enhancing the exploration capabilities algorithms may struggle due to the sheer size of the search space or the
and convergence speed. This approach leverages the superposition and need for rapid convergence46. In addition, the evolution of quantum
entanglement of quantum information (i.e., vectors in Hibbert space), chromosomes can save memories by avoiding maintaining groups of
enabling the algorithm to explore multiple solutions simultaneously and binary vectors in CGA, which represent the population in the gen-
avoid local optima more effectively. By combining these quantum principles eration. QGA also utilizes the concept of quantum entanglement,
with the adaptive search capabilities of genetic algorithms, QGA can provide enabling a more effective exchange of information between solutions.
a more efficient and scalable solution for complex optimization problems. This makes QGA more adept at avoiding premature convergence to
The iterative process of QGA, which includes quantum-inspired mutation local optima, which is a common pitfall in CGA. In the evolution of
and crossover operations, maximizes a comprehensive search of the solu- QGA, the quality of the quantum chromosome is evaluated by calcu-
tion space, leading to higher precision and faster convergence compared to lating the fitness values of independent measurements of the quantum
traditional methods28. chromosome (the details of quantum measurement are in the Method
The Planar multilayer (PML) system is very widely used in opto- section). The fitness values are evaluated by a figure-of-merit (FOM),
electronics, photonic crystals, and anti-reflective coatings due to their which is used to describe how close between a designed PML structure
simplicity in manufacturing and flexibility in design29. Recently, it has been and the ideal TRC. The FOM can be calculated by:
used for the design of transparent radiative coolers (TRC)30–32, which block
R λ2  2
non-visible solar light and prevent excessive heating of rooms or 10 TlðλÞ  Tlideal ðλÞ dλ
λ1
compartments33–36 without external energy29,37–39. In general, PML structures FOM ¼ R λ2 ð1Þ
λ SðλÞ dλ
2
consist of dielectric materials, metals or organic thin films with distinctive 1

refractive indices, and a spectral response across the solar wavelength range
can be determined by the special combination and order of the material (the details of the FOM and parameters definition are in the Supple-
layers in the PML structure. The optimization for PML can be formulated mentary Information, Section 1). Minimizing the FOM will allow the
into a combinatorial optimization problem for the stack of different can- designed TRC to approach the ideal TRC, so that the fitness value in the
didate materials40. As the number of material layers increases, the dimension QGA is defined as f = −100 × FOM. The measurements with high fitness
of the optimization problem will rise to a level that is prohibitive for many values provide genetic information for the evolution direction of quantum
traditional methods, and there will be lots of local optima present in the chromosomes.
objective function. Researchers have proposed a variety of optimization Compared with the CGA, QGA employs the probability amplitudes of
strategies for exploring potential optimal PML structures, such as needle qubits to encode chromosomes (as depicted in Eq. 12) and utilizes quantum
optimization29,40,41, memetic method42,43 and deep reinforcement learning44. rotation gates to execute the chromosomal updated operations. Conse-
However, these methods often encounter challenges with converging to quently, the formulation of the quantum rotation gates serves as a critical
local minima. Therefore, an advanced algorithm with exceptional good aspect of QGA, directly influencing the algorithm’s performance. Here, we
global search capabilities is needed. Kitai et al. 45 and Kim et al. 24 utilized apply rotation-Y (Ry) gates on each qubit in the chromosome to realize the
quantum annealing (QA) and factorization machine (FM) for the design of evolution. The Ry gate can be defined in a matrix form as28,47,48:
PML. These algorithms, based on quantum annealing, deliver a superior " #
optimization speed and precision compared to classical computers. cos θ2  sin θ2
In this work, we introduce an optimization algorithm based on a QGA RyðθÞ ¼ ð2Þ
sin θ2 cos θ2
for binary combinatorial optimization problems. As a test case, we use it to
design PML for TRC applications. This optimization algorithm can utilize
general surrogate models and find the optimum by QGA. We choose the The updated process is:
Random Forest (RF) as the surrogate model in this study. Numerical 2 3
experiments are conducted for structures with 6 to 20 layers, and the results " #   θt θtk  t
atþ1  t  atk cos 2k  sin a
show that our QGA-facilitated optimization algorithm can converge to k
¼ Ry θk ¼ 4 2 5 k ð3Þ
comparable solutions as QA and overperforms classical genetic algorithm btþ1
k btk θt
sin k cos
θtk btk
2 2
(CGA) on both convergence speed and global search capability. Further-
more, due to the advantages brought by the RF model, fewer iterations are k ; bk 
where ½αtþ1 tþ1 T
and ½αtk ; btk T
represent the probability amplitudes of the
needed for the QGA to converge to the optimal solution compared with k-th qubit in the chromosome at the t + 1-th generation and the t-th gen-
QA-facilitated optimization. The proposed method provides a powerful tool eration, respectively. θtk is the rotating angle for the Ry gate. We used an

npj Computational Materials | (2024)10:257 2

[Link] Article

Table 1 | Adjustment Strategy of Rotating Angle initialized to a uniform superposed state by the Hadamard gate as:

xi* besti f(x) > θ s(ai, bi) ***     " p1ffiffi #

f(best) ** 1 1 1 1 2
aibi > 0 aibi < 0 ai = 0 bi = 0 H ¼ pffiffiffi ;H ¼ 1 ð7Þ
2 1 1 0 pffiffi
0 0 False 0 0 0 0 0 2

0 0 True 0 0 0 0 0
Then, a series of Ry gates with randomly generated angles is applied to
0 1 False θi +1 −1 0 ±1
each qubit to produce a random quantum chromosome for QGA. Except
0 1 True θi −1 +1 ±1 0 for the first QGA iteration, other iterations will start with the initialized
1 0 False θi −1 +1 ±1 0 quantum chromosome which also incorporates the solution of former QGA
1 0 True θi +1 −1 0 ±1 iteration with weighing factor wp as shown in Eq. 6. The determination of wp
1 1 False 0 0 0 0 0
is based on the following formula:
1 1 True 0 0 0 0 0
1 best current
* xi and besti represent the i-th bit of chromosomes and best individual, respectively. wp ¼ ð8Þ
2 best all
** f(·) is the fitness evaluation function. The best individual has the highest fitness.
*** s(ai, bi) is the direction of the rotating operation. +1 represents clockwise rotation.
where bestall and bestcurrent are the best fitness values of all time and in the
current iteration, respectively. If the current iteration successfully converges
to or exceed the fitness value of all time, bestall will equal to bestcurrent at the
adaptive adjustment strategy28 for the quantum rotation angle so that the end of the iteration, so that wp = 0.5. Otherwise, bestall will be higher than
direction and magnitude are updated dynamically during the evolutionary bestcurrent at the end of the iteration, which means that the initialization of the
process. A large rotating angle is set early in the evolutionary process to next iteration will have higher degree of randomness to encourage
quickly explore the solution space and find regions likely to contain optimal exploration of the entire landscape.
values. To accurately find the optimal value, we reduce the magnitude of Additionally, to further mitigate the risk of the algorithm falling into a
rotating angle as evolution continues so that it eventually converges to a local optimum, we introduce an operation which we call memory corrup-
global optimum. The update of rotating angle follows the following formula: tion. This operation allows the optimal result, inherited from the previous
  cycle, to be discarded with a certain probability. This ensures that the
θmax  θmin incoming iteration does not simply inherit the optimal solution from the
θi ¼ θmax  ×i ð4Þ
N previous one. Consequently, an observable decay of the best solution can
occur during the evolution of QGA, preventing stagnation and encouraging
where θi is the value of the rotating angle of the i-th generation, N is the max the search for new potential solutions.
generation, θmax and θmin is the upper bound and lower bound of θ. The The evolution of the QGA will terminate at a pre-defined maximum
adjustment strategy of rotating angle is shown in Table 149. generation. However, if the evolution remains stagnant for a long time, it will
In the CGA, the crossover operation in the evolution allows for an be considered to have reached a local optimum and thus will be terminated.
expansive exploration of the solution space. However, the quantum chro- The termination criteria can vary based on the specific optimization pro-
mosome itself has the property of individual diversity resulting from blem. For the optimization problem discussed in this paper, we have
quantum superposition. So, there is no need to perform the crossover determined that if the QGA evolution remains in a certain state for more
operation in the QGA. On the other hand, mutation is the operation that can than half of the maximum generation, the evolution will be terminated
ensure sufficient variety in the population to avoid local optima, which is automatically.
important for both CGA and QGA. Different from CGA, quantum muta- In summary, the QGA process is described in Fig. 1.
tion will appear on the quantum chromosome and completely reverse the
individual’s evolutionary direction by swapping the value of probability Optimization of PML TRC
amplitudes a and b of mutated qubits, which is implemented by X-Gate on We consider the design of an N layer PML for TRC (as shown in Fig. 2a).
the randomly selected qubits as47: The thickness of designed PML is 1200 nm in total. Each layer of the PML
structure can be one of four candidate materials: silicon dioxide (SiO2),
   t " t# silicon nitride (Si3N4), titanium dioxide (TiO2), or aluminum oxide (Al2O3),
0 1 a bk
X¼ ; X tk ¼ ð5Þ and each material is assigned one of the combinations of two binary labels
1 0 bk atk
(00 = SiO2, 01 = Si3N4, 10 = TiO2, 11 = Al2O3). The concatenation of these
binary labels in order gives a 2N-long vector representing the structure. The
Additionally, at the start of each QGA iteration, the genetic informa- optimization of PML structures is to minimize the FOM. The perfect PML
tion of the optimal individual obtained from the previous QGA iteration will can block all UV and IR light while allowing all visible light to transmit
be partially incorporated into the initialization of the quantum chromo- through. The proposed QGA, as an optimizer, is utilized to perform the
some, which reinforces the retention and utilization of valuable genetic optimization. As a comparison, we have also used CGA and QA for the same
information, leading to more efficient convergence towards optimal solu- optimization problem. The details of QA can be found in Supplementary
tions. Therefore, the initialization of the quantum chromosome q in itera- Information, Section 3.
tion τ is: Fig. 2(b) shows the schematic of the QGA-facilitated optimization
" " 1 # #
algorithm proposed for PML optimization. The algorithm is based on active
  pffiffi learning scheme with iterations of random forest (RF) training, QGA
2
qðτ ¼ 0Þ ¼ Ry θrand  for k ¼ 1; 2; . . . ; n
k p1ffiffi optimization, and TMM calculations. The details of the algorithm can be
"  it1
2 k
" 1 #
pffiffi
# ð6Þ found in Method section.
ak   2
qðτ > 0Þ ¼ wp 
bk
þ ð1  wp Þ  Ry θrand k  1 for k ¼ 1; 2; . . . ; n
pffiffi
We first take N = 6 as a benchmark study to test our algorithm, for
2 k
which we can afford performing an exhaustive search by calculating the
where θrand is the randomly generated initial angle between 0 to π, wp is the optical properties using the transfer matrix method (TMM), which is the
weighing factor to balance the genetic information from prior QGA itera- most efficient method to calculate optical characteristics of PML structures
tion and random initialization in the current iteration. Each qubit is firstly and evaluate the FOM of every one of all the 46 = 4096 possible structures.

npj Computational Materials | (2024)10:257 3

[Link] Article

(a)

UV
IR binary vector representation
Solar Spectrum 10 01 11 00 … 00 01 10 11

0 0 = SiO2
0 1 = Si3N4
1 0 = TiO2
1 1 = Al2O3
SiO2

Visible Photons

(b)

Fig. 2 | Schematics of PML TRC structure design. a The schematic structure of

the PML in the TRC. With proper binary embedding for candidate materials, the
PML can be mapped into a binary vector. b The workflow of active learning
iteration between RF, QGA and TMM.

Fig. 1 | The Workflow of QGA. A quantum chromosome is first initialized, followed

by the measurement and fitness evaluation of measured individuals. The best
individual is selected from all the measurements, and its genetic information is used
to guide the evolution of the quantum chromosome. The process repeats until the
termination conditions are satisfied.

The optimal structure, denoted as [10 10 00 10 01 11], is identified with a

FOM of 1.7713 from this brute force exhaustive search. We then used our
QGA-facilitated optimization to solve the same N = 6 problem. The QGA-
facilitated optimization is implemented with an initial training set with the
amount of data of m = 25 and a maximum iteration number of 10. In each
generation, the quantum chromosome undergoes 25 measurements, and
fitness evaluations are performed on these measurements using the RF sur-
rogate model to discern the evolution direction of the quantum chromosome.
An adaptive rotating angle, as defined in Eq. 4, decays from the upper to the
lower bounds of 0.1π and 0.01π, respectively. Throughout the evolution
process, the mutation rate is maintained at 0.001. The QGA terminates after
100 generations or if it reaches a converged state before that. Fig. 3 shows the
evolution of FOM in the QGA optimization. The blue curve, orange curve Fig. 3 | The evolution of FOM for N = 6 in QGA-facilitated active learning
and green curve represent the best FOM in the current generation, the average optimization. The blue curve, orange curve and green curve represent the best FOM
FOM in the current generation and the best FOM of all time, respectively. The in the current generation, the average FOM in the current generation and the best
x-axis is the step in the entire optimization, and each iteration of the active FOM of all time, respectively. In each iteration of our active learning scheme, there
learning has 100 steps (correspond to 100 generations in QGA). As the figure are 100 generations for the QGA.
shows, in each iteration, the FOM decays to a converged value, and the next
iteration starts with a newly initialized quantum chromosome which partially
include the information from the last iteration (the details of the quantum optima. This benchmark test indicates the effectiveness and efficiency of our
chromosomes initialization is shown by Eq. 6). After only 5 iterations, the QGA-facilitated active learning optimization scheme.
FOM has successfully converged to 1.7713, which is consistent with the The optimization is subsequently carried out for structures with N = 8,
solution obtained from the exhaustive search. Compared with the 4096 TMM 10, 16, and 20. Optimizing structures with higher N demands a larger
calculations required by the brute force exhaustive search, QGA only requires population size and an increased number of generations in QGA. In addi-
~125 TMM calculations (details in the Methods section) in 5 active learning tion, a larger initial dataset may also be needed to start the iterations effi-
iterations to label the solution of each iteration before getting the global ciently. The parameters for the optimization are detailed in the

npj Computational Materials | (2024)10:257 4

[Link] Article

Fig. 4 | Evolution of the best FOM in QGA

and CGA. Panel (a) ~ (d) represent the results of N =
8, 10, 16 and 20, respectively. The solid lines repre-
sent the average values from 5 multiple independent
numerical experiments. The shadow regions repre-
sent the range of the best FOM from different trials.
The horizontal green dash lines in the plots are the
ground truth from exhaustive search (for N = 8 and
10) or the best solution from QA-assisted optimi-
zation algorithm (for N = 16 and 20).

Supplementary Information, section 2. Fig. 4 shows the evolution of the best

FOM of all time for N = 8, 10, 16 and 20, and the long-term convergence
states in the high dimensional cases are not shown to highlight the FOM
evolution at the early stage. For low dimensional cases (N = 8 and 10), our
algorithm can discover accurate ground state obtained from exhaustive
search faster than CGA-facilitated optimizations. For higher dimensional
problems (such as N = 16), our QGA algorithm can converge to the same
structure as QA-facilitated optimization (using a real quantum annealer, D-
Wave) within 7 iterations (under 5000 generations in total), while CGA start
to fail obtaining a comparable result with QA. For the problem with N = 20,
both CGA and QGA algorithms cannot find the structure that is as good as
QA finds. That is because the extremely large dimension of the search space
requires an extremely large number of measurements in each generation to
explore, which is prohibitively computationally expensive using a classical
computer. Theoretically, the randomness introduced by quantum mea-
surements, given a sufficient number of measurements, will guide the
algorithm to explore the search space. The process involved in finding a local
optimum, escaping from it, and then identifying a better solution, ultimately Fig. 5 | Evolution of the best FOM in QGA for 16-layer PML TRC with different
converging to a global optimum or a proper local optimum, can be very thicknesses (800 nm, 1000 nm and 1200 nm). The solid lines and shadow represent
computationally intensive and time-consuming when using a classical the average values and the range of the best FOM from 5 multiple independent numerical
computer to mimic quantum operations50. Especially when the dimension experiments. Dash lines represent the results from QA-assisted optimization algorithm.
of the problem increases, the required measurement of quantum chromo-
somes and the fitness evaluation of each measurement result will be the
bottleneck of computational efficiency and will also be the decisive factor in important for TRC performance. We explore how varying the total thick-
whether the algorithm can converge to the optimal solution. To improve the ness of the PML structure affects its performance. Specifically, 16-layer PML
performance of the algorithm on higher-dimensional problems, future work structures with different overall thicknesses (1200 nm, 1000 nm, and
can focus on accelerating the measurement of quantum states, so that most 800 nm) are optimized using the proposed algorithm. Fig. 5 shows the
of the evolution operations of QGA can be migrated to real quantum evolution of the best FOM in QGA for the three different PML TRCs. QA-
devices, greatly reducing the amount of computation undertaken by clas- assisted optimizations are used to benchmark optimized results for QGA-
sical computers during the optimization, thereby improving the ability of facilitated optimization. For all the three systems, out QGA-facilitated
QGA algorithm on high-dimensional problems. Nevertheless, due to the algorithm can still converge to the same solution as QA-facilitated algo-
inherent quantum properties, QGA still outperforms CGA on both con- rithm. The effect of PML thickness on TRC performance is not monotonic.
verging speed and exploring ability. Compared to the 1200 nm structure, both the 1000 nm and 800 nm struc-
The thicknesses of all the PML TRCs discussed above are fixed at tures show better performance, with the 1000 nm structure outperforming
1200 nm. However, the overall thickness of the PML structures is also the 800 nm structure. This makes thickness a key variable in the design of

npj Computational Materials | (2024)10:257 5

[Link] Article

TRC materials. However, since thickness is a continuous variable, it cannot Computational efficiency analysis
be directly optimized using QA. Therefore, developing an efficient dis- One crucial metric for evaluating the efficiency of an algorithm is its com-
cretization method for continuous variables such as thickness will become putational cost. In the context of the PML optimization using active
an important factor in applying quantum algorithms to material design. In learning, the total number of TMM calculations, which is the rate-limiting
future work, we will focus on developing discretization methods to enable step in the iteration, serves as a key indicator of this cost. When optimizing
the application of quantum algorithms for material design. an N-layered structure, an exhaustive search approach would necessitate 4 N
TMM calculations to explore all possible solutions before identifying the
true global optimum.
Fig. 6 shows the relationship between the number of required fitness
evaluations by TMM calculations and the number of layers in the design by
1e+14
Exhaustive Search using exhaustive search, QGA and QA. For QGA, only the cases that found
1e+12 QGA max (only success cases) the global optima are counted in comparison. Note, the y-axis is plotted on a
QA log10 scale. The efficacy of our QGA-facilitated and QA-facilitated opti-
mization is highlighted by their substantially reduced need for TMM cal-
TMM Calculations

1e+10
culations as compared to exhaustive search methods. This advantage is
1e+08 further amplified by the high-performance nature of the RF model, which
enables the max required TMM calculation by the QGA-facilitated opti-
1e+06 mization algorithm basically the same order of magnitude as the QA-
assisted optimization algorithm, but during the optimizations, the QGA-
1e+04 facilitated algorithm requires even fewer TMM calculations as the it gra-
dually converges to the optimal structure and producing repeated solutions
1e+02 in the labeling. This results in a significant alleviation of the burden tradi-
tionally associated with data acquisition, which is the TMM calculation in
1 our case, but can be other physics-based modeling or experiments in other
6 8 10 12 14 16 18 20 optimization tasks. This advantage is believed to be from the accuracy of the
N surrogate model, which is further discussed in the next section.
Fig. 6 | Comparison of computational cost for exhaustive search, QGA and QA. Comparison between random forest (RF) and factorization
For QGA, we only take the cases that successfully converge with global optima (from
machine (FM) model
exhaustive search for low dimensional cases) and QA results (for high dimensional
Although the QGA’s global search capability is currently not as good as QA
cases). The numbers of TMM calculations in QGA-facilitated optimization are less
than the max evaluation, which is the population size × iteration number.
in the high dimension optimization cases, the superiority of the RF model
over factorization machine (FM), which is the required surrogate for QA as

Fig. 7 | Comparison between FM and RF models (a) N=8 (b) N = 10

prediction RMSE. Panel (a) ~ (d) show the com- 3 3
parisons for (a) N = 8, (b) N = 10, (c) N = 16, and (d) FM FM
N = 20. RMSE is evaluated on the same test dataset 2.5 RF 2.5 RF
for FM and RF models. The error bars are calculated
from the results of 10 independent experiments.
2 2
RMSE

RMSE

1.5 1.5

1 1

0.5 0.5

0 0
25 50 75 100 25 50 75 100
Num of Data Num of Data

(c) N = 16 (d) N = 20
3 3
FM FM
2.5 RF 2.5 RF

2 2
RMSE

RMSE

1.5 1.5

1 1

0.5 0.5

0 0
25 50 75 100 25 50 75 100
Num of Data Num of Data

npj Computational Materials | (2024)10:257 6

[Link] Article

Fig. 8 | QGA evolution with RF surrogate model

and FM surrogate model. a N = 8 and (b) N = 16.
Solid lines and shadows represent average FOM
evolution and range of multiple trials.

it can be mapped into a QUBO formulation, affords our proposed QGA- Towards More Applications – Optimization of Optical Diodes
facilitated optimization algorithm a more reliable searching landscape for In addition to the application in the TRC design, our QGA-facilitated
high-dimensional problems. Fig. 7 provides a comparative analysis of the optimization algorithm can be applied to other material design problems.
Rooted Mean Squared Error (RMSE) for FM and RF. Both are trained using Here, we test the performance of our algorithm on designing metamaterials
TMM-calculated FOM data from of 8-layer, 10-layer, 16-layer, and 20-layer optical diodes. Nanophotonic structures that exhibit asymmetrical power
structures, with varying numbers of data points (25, 50, 75, and 100) ran- (intensity) transmission can be considered optical diodes, following the
domly selected. 10 independent experiments have been performed and the concept of electrical or thermal diodes that have asymmetrical transport
standard errors of the mean are calculated as the error bars. Training is properties52–54. An optical diode permits the intensity of light to transmit in
implemented with 5-fold cross-validation and Stochastic Gradient Descent one direction (i.e., forward direction) but blocks it in the reverse direction
(SGD), and RMSE is evaluated on a randomly selected test set. The results (i.e., backward direction). The FOM of an optical diode can be defined as the
show that RF models outperform FM on the test sets. Moreover, as the data difference between the forward (TF) and backward (TB) transmissivities (i.e.,
dimensionality increases, the advantages of RF become more apparent. FOM = TF - TB)55, and a larger FOM indicates the better performance of the
From Fig. 7, especially in panels (c) and (d), it can be seen that the RF model optical diode.
trained with less data can achieve higher accuracy than the FM model The stratified volume diffractive film composed of metal and dielectric
trained with more data, which indicates that the RF model requires less data materials can be considered as a system for optical diodes. Fig. 10a shows the
to achieve higher prediction accuracy than FM model, which should be the schematic structure of the thin-film optical diode with a stratified volume
reason why fewer iterations are necessary in our QGA-facilitated optimi- diffractive film, in which it has a unit cell consisting of metallic gratings and a
zation algorithm than the QA-assisted optimization. dielectric spacer. Fig. 10b shows the configuration of the unit cell, which is
To further confirm this hypothesis, the trained RF and FM discretized with rectangular pixels, as proposed by Kim et al. in ref. [55]. The
models are subsequently employed as surrogate models for the QGA material selection of a pixel can be encoded into a binary digit as “0” for a
in optimizing PML structures with dimensions of N = 8 and 16. Fig. 8 dielectric or “1” for a metal medium. The pixelated metamaterial structure is
illustrates the evolution curves of QGA when utilizing RF and FM as then represented by a binary vector with a length of N. FOMs of the
surrogate models with the same initialization. It can be observed that structures are calculated using the rigorous coupled-wave analysis (RCWA)
for the low-dimensional problem of N = 8, QGA always converges to method56. Here, we test our algorithm on two cases: wavelength λ = 600 nm
the optimal structure regardless of the surrogate model employed, (N = 40) and 800 nm (N = 32). Fig. 10c shows the evolution of the best FOM
and the converging speed using RF is faster than FM. However, for as a function of optimization steps for both cases. For these high dimen-
the higher-dimensional scenario (N = 16), employing RF as the sur- sional optimization problems, although still limited by computational
rogate model enhances the speed of QGA’s convergence greatly. This resources to find the same optimal structures as QA-assisted optimization
substantiates the advantage of using RF as a more general surrogate algorithm found (0.8064 for λ = 600 nm, 0.8648 for λ = 800 nm), our
model over FM. algorithm can still lead the exploration of the design space towards the right
direction within 50 iterations (the population sizes are 200 in both cases).
Energy Saving Analysis The number of RCWA calculations in the QGA-facilitated algorithm is
Lastly, we estimated the benefit of the 16-layer and 20-layer TRC designed under 10,000 in both cases, which are in the same order of magnitude as the
by our algorithm as a potential window material by calculating the energy it thousands of RCWA calculations needed for the QA-assisted algorithm.
can save annually in the U.S. using EnergyPlus24,51. The annual energy saving Comparing with CGA-facilitated optimization, our QGA-facilitated algo-
over the surveyed U.S. cities is shown in Fig. 9a, b. On average, the appli- rithm can still converge faster to better optimal structures, which has been
cation of the 16-layer TRC as a window material would yield an annual already been shown in the TRC case.
energy saving of 33.58 MJ/m2 over the surveyed locations. Even for the 20-
layer TRC structure, despite not reaching the global minimum, can still Discussion
contribute to an average annual energy saving of 26.03 MJ/m2. Fig. 9c, d In this work, we introduce an optimization algorithm based on QGA
show the energy savings for the top 15 energy consuming states in the U.S. to address the challenge of finding the optimal in binary combina-
For the hot states (e.g., Arizona, Nevada, and Hawaii), our designed 16-layer torial problems. We used PML for TRC as an application example.
TRC can potentially save ~30% of the cooling energy compared to con- The combination of QGA and RF regression model iteratively
ventional windows. The same calculations have also been performed for 20- operates in an active learning framework to obtain global optimum in
layer TRC, and an energy saving over 20% can be achieved. Fig. 9e shows the complex and discrete search spaces. QGA excels in aspects like
comparison between the transmitted irradiance through the ideal and our smaller population size, faster convergence speed, superior global
optimized 16-layer and 20-layer TRCs. The transmitted irradiance of the 16- search ability, and robustness, greatly outperforming conventional
layer structure has better alignment with the ideal TRC, showing better genetic algorithms. Moreover, compared to employing FM (QUBO)
performance than the 20-layer structure. to describe the search space, the RF regression model demonstrates

npj Computational Materials | (2024)10:257 7

[Link] Article

(a) (b)

(c) (d)

(e) 2
nm-1 )

Ideal TRC
Designed 16-layer TRC
-2

Designed 20-layer TRC

1.5
Transmitted Irradiance (W m

0.5

0
500 1,000 1,500 2,000 2,500
Wavelength (nm)

Fig. 9 | Energy saving analysis for design PML TRC structures. Panel (a) and (b) the 20-layer TRC designed by our QGA-facilitated optimization. e The transmitted
show the estimated cooling energy saving across the U.S. by using the 16-layer and irradiance through the target ideal and our optimized 16-layer and 20-layer TRCs by
20-layer TRC as the window material, respectively. The annual energy savings of the QGA-facilitated active learning scheme.
top 15 energy-consuming states over U.S. by applying (c) the 16-layer TRC and (d)

stronger reliability in high-dimensional problems. Hence, unless the systems exist in definite states, quantum computing operates on the prin-
dimensionality of the problem is unreachable for classical computers, ciples of superposition and entanglement, allowing the system to exist in
QGA can be more efficient compared QA not to mention CGA. multiple states simultaneously57–59. The state of a quantum system is
Furthermore, any predictive model that is well-suited for accepting described by a probabilistic wave function, the square of which provides the
binary vectors as input will be accessible for our QGA-based opti- probabilities for the possible states28. This attribute expedites computation in
mization framework, and this will provide more freedom when quantum systems by order of magnitude compared to classical systems.
compared with QA-based optimization. At the same time, the results The evolution of qubits through so-called quantum channels is carried
in this work point out the bottleneck of current QA-facilitated out by unitary quantum gates that manipulate the qubits just like classical
optimization scheme and strongly indicate the necessity for devel- bits are manipulated by logic gates in a computer60,61. These unique char-
oping FM that with 3rd or higher orders and accessible by quantum acteristics of quantum computing – superposition, entanglement, and
annealers. quantum gate operations – confer it with unparalleled computational
power62. Integrating quantum principles into optimization algorithms offers
Methods potential improvements, enhancing traditional methods and aiding in
Quantum computing complex problem-solving.
Quantum computing, an intriguing application of quantum mechanics in The basic unit for information storage is qubit in the quantum com-
algorithmic computation, significantly diverges from classical computing puter. Unlike classical bits that can be in a state of either 0 or 1, a qubit can
primarily due to its intrinsic parallelism. Unlike classical computing where exist in a state of superposition, representing 0 and 1 simultaneously. The

npj Computational Materials | (2024)10:257 8

[Link] Article

Fig. 10 | Schematics of optical diode design. a the

thin-film optical diode with a stratified volume dif-
fractive film for optimization and (b) the pixelated
unit cell of the structure. c Evolution of FOM in
QGA-facilitated algorithm for 600 nm and 800 nm
wavelength. The dimensions of the binary vector are
N = 40 and N = 32, respectively. The average FOM
evolutions of the two cases in CGA-facilitated
optimizations are compared with that in QGA-
facilitated optimizations.

super position state of a qubit can be expressed as follows: Active learning scheme with RF and QGA
We employ an optimization algorithm based on the active learning
jφi ¼ aj0i þ bj1i ð9Þ framework. To circumvent the computationally intensive TMM cal-
culation, we implement a RF as a surrogate model, significantly
where, |φ〉 is the superposition state. a and b are complex numbers which accelerating the calculation speed12,70 of the FOM per instance. Initially,
denote the probability amplitudes of the corresponding ground state |0〉 or the RF model is trained on m data points calculated by the TMM. Due
|1〉, and |a | ² + |b | ² = 1. Therefore, an n-qubits quantum register can store to the limited quantity of training data, the accuracy of the Random
the quantum state which is the coherent superposition of 2n ground states: Forest is not expected to be high initially or whole optimization space.
Therefore, an iterative process between QGA and RF training is
X
2n implemented within the active learning loop. During each iteration, the
jϕi ¼ ci jϕi i ð10Þ FOM of individuals in the QGA is evaluated using RF. The FOM cal-
i¼1
culated by RF surrogate model is mapped to the fitness value in the
where ci are the probability amplitudes that satisfying: QGA, and all the optimal structures found in each QGA generation
having lower FOMs than the best structure have found before will be
X
2n recorded and the latest n solutions will be put into the TMM calcula-
jci j2 ¼ 1 ð11Þ tions to label true FOM, which added to the database to train a new RF
i¼1
model, thereby promoting the progression of iterations. The number of
If we use a binary vector to encode qubits on the polymorphic problem, labeled optimal structures after each QGA evolution, n, is set to be the
the system with n qubits can be expressed as follows63: same as population size. Therefore, the max number of TMM calcu-
lations in the QGA-facilitated algorithm is total iteration numbers to
 
a1 a2 a3 an get converged × the population size in each QGA generation.
q¼  ð12Þ In the algorithm, the structures of the TRC are represented by a series of
b1 b2 b3 bn
binary vectors, which are also the input of TMM and RF calculations. After
where |ak | ² + |bk | ² = 1 (k = 1, 2, …, n), and each pair of probability training the RF model, the QGA starts with the quantum chromosome
amplitudes [ak, bk]T represent a qubit in the system. When the quantum initialized according to Eq. 6, and the fitness evaluation in the QGA evo-
system is measured, coherence will disappear, and the quantum system will lution is implemented by the surrogated RF model. After obtaining the best
collapse to a definite state |Φi〉 with probability given by the squared mag- solution from QGA, TMM is called to verify the FOM of the best solution
nitude of the respective probability amplitude |ci | ² according to Eq. 10. The and add the data to the training set to retrain the RF model. If the solution of
probability of qubit k being measured in state |0〉 (or |1〉) will be |ak | ² (or QGA has been included in the training set, a randomly selected structure
|bk | ²) as Eq. 12 described. and its FOM is added to the training set to further diversify the applicability
of the RF model in the optimization space. With the iterations ongoing, the
Random forest algorithm accuracy of RF will be improved in the design space, by which the fitness
Decision trees, particularly the Classification and Regression Tree (CART) evaluation will be more reliable for finding the best solution.
method, are known for their simplicity and interpretability in supervised
learning64,65. To counter the potential overfitting problem of single decision Data availability
tree, the Random Forest (RF) algorithm, which was proposed by Breiman Requests for data and materials should be sent to the corresponding authors
et al. 66 in 2001, is always employed as an ensemble method to enhance or Z.X. (zxu8@[Link]).
model robustness and predictive accuracy. The RF algorithm operates by
constructing a multitude of decision trees and averaging related predictions Code availability
from individual trees, and therefore it will yield a more accurate final pre- The underlying codes for this study are available from the corresponding
diction than any individual tree could offer67–69. author or Z.X. (zxu8@[Link]) upon reasonable request.

npj Computational Materials | (2024)10:257 9

[Link] Article

Received: 21 February 2024; Accepted: 24 October 2024; 24. Kim, S. et al. High-Performance Transparent Radiative Cooler
Designed by Quantum Computing. ACS Energy Lett. 7, 4134–4141
(2022).
References 25. Narayanan, A. & Moore, M. in Proceedings of IEEE International
1. Pope, M. A. & Aksay, I. A. Structural design of cathodes for Li‐S Conference on Evolutionary Computation. 61-66.
batteries. Adv. Energy Mater. 5, 1500124 (2015). 26. Han, K.-H. & Kim, J.-H. Genetic Quantum Algorithm and its Application
2. Zhang, X. et al. Multiscale understanding and architecture design of to Combinatorial Optimization Problem. 1354–1360 (2003).
high energy/power lithium‐ion battery electrodes. Adv. Energy Mater. 27. Lv, Y. & Li, D. in 2008 Fourth International Conference on Natural
11, 2000808 (2021). Computation. 460–464 (IEEE).
3. Fayaz, H. et al. Optimization of thermal and structural design in 28. Wang, H., Liu, J., Zhi, J. & Fu, C. The Improvement of Quantum
lithium-ion batteries to obtain energy efficient battery thermal Genetic Algorithm and Its Application on Function Optimization. Math.
management system (BTMS): a critical review. Arch. Comput. Probl. Eng. 2013, 730749 (2013).
Methods Eng. 29, 129–194 (2022). 29. Raman, A. P., Anoma, M. A., Zhu, L., Rephaeli, E. & Fan, S. Passive
4. Liu, D., Tan, Y., Khoram, E. & Yu, Z. Training Deep Neural Networks for radiative cooling below ambient air temperature under direct sunlight.
the Inverse Design of Nanophotonic Structures. ACS Photonics 5, Nature 515, 540–544 (2014).
1365–1369 (2018). 30. Li, W., Li, Y. & Shah, K. W. A materials perspective on radiative cooling
5. Liu, Z., Zhu, D., Rodrigues, S. P., Lee, K.-T. & Cai, W. Generative structures for buildings. Sol. Energy 207, 247–269 (2020).
Model for the Inverse Design of Metasurfaces. Nano Lett. 18, 31. Lee, M. et al. Photonic structures in radiative cooling. Light.: Sci. Appl.
6570–6576 (2018). 12, 134 (2023).
6. Ma, W., Cheng, F. & Liu, Y. Deep-Learning-Enabled On-Demand 32. Vall, S., Medrano, M., Solé, C. & Castell, A. Combined Radiative
Design of Chiral Metamaterials. ACS Nano 12, 6326–6334 (2018). Cooling and Solar Thermal Collection: Experimental Proof of
7. Kudyshev, Z. A., Kildishev, A. V., Shalaev, V. M. & Boltasseva, A. Concept. Energies 13, 893 (2020).
Machine learning–assisted global optimization of photonic devices. 33. Arrés Chillón, J., Paulillo, B., Mazumder, P. & Pruneri, V. Transparent
Nanophotonics 10, 371–383 (2021). Glass Surfaces with Silica Nanopillars for Radiative Cooling. ACS
8. So, S. & Rho, J. Designing nanophotonic structures using conditional Appl. Nano Mater. 5, 17606–17612 (2022).
deep convolutional generative adversarial networks. Nanophotonics 34. Lei, M.-Q. et al. Transparent radiative cooling films containing
8, 1255–1261 (2019). poly(methylmethacrylate), silica, and silver. Optical Mater. 122,
9. Liu, Z., Zhu, Z. & Cai, W. Topological encoding method for data-driven 111651 (2021).
photonics inverse design. Opt. Express 28, 4825–4835 (2020). 35. Jaramillo-Fernandez, J. et al. Highly-Scattering Cellulose-Based
10. Ma, W., Cheng, F., Xu, Y. & Wen, Q. & Liu, Y. Probabilistic Films for Radiative Cooling. Adv. Sci. 9, 2104758 (2022).
Representation and Inverse Design of Metamaterials Based on a Deep 36. Jin, Y., Jeong, Y. & Yu, K. Infrared‐reflective transparent hyperbolic
Generative Model with Semi-Supervised Learning Strategy. Adv. metamaterials for use in radiative cooling windows. Adv. Func. Mater.
Mater. 31, 1901111 (2019). 33, 2207940 (2022).
11. Ma, W. & Liu, Y. A data-efficient self-supervised deep learning model 37. Chen, Z., Zhu, L., Raman, A. & Fan, S. Radiative cooling to deep sub-
for design and characterization of nanophotonic structures. Sci. China freezing temperatures through a 24-h day–night cycle. Nat. Commun.
Phys., Mech. Astron. 63, 284212 (2020). 7, 13729 (2016).
12. Wiecha, P. R., Arbouet, A., Girard, C. & Muskens, O. L. Deep learning 38. Zhai, Y. et al. Scalable-manufactured randomized glass-polymer
in nano-photonics: inverse design and beyond. Photon. Res. 9, hybrid metamaterial for daytime radiative cooling. Science 355,
B182–B200 (2021). 1062–1066 (2017).
13. Deng, L., Xu, Y. & Liu, Y. Hybrid inverse design of photonic structures 39. Li, T. et al. A radiative cooling structural material. Science 364,
by combining optimization methods with neural networks. Photonics 760–763 (2019).
Nanostruct. - Fundamentals Appl. 52, 101073 (2022). 40. Tikhonravov, A. V., Trubetskov, M. K. & DeBell, G. W. Application of
14. Ma, W. et al. Deep learning for the design of photonic structures. Nat. the needle optimization technique to the design of optical coatings.
Photonics 15, 77–90 (2021). Appl. Opt. 35, 5493–5508 (1996).
15. Mahrous, H. et al. A Multi-Objective Genetic Algorithm Approach for 41. Tikhonravov, A. V. Some theoretical aspects of thin-film optics and
Silicon Photonics Design. Photonics 11, 80 (2024). their applications. Appl. Opt. 32, 5417–5426 (1993).
16. Ren, Y. et al. Genetic-algorithm-based deep neural networks for 42. Li, W., Shi, Y., Chen, Z. & Fan, S. Photonic thermal management of
highly efficient photonic device design. Photon. Res. 9, B247–B252 coloured objects. Nat. Commun. 9, 4240 (2018).
(2021). 43. Shi, Y., Li, W., Raman, A. & Fan, S. Optimization of Multilayer Optical
17. Cai, H., Sun, Y., Wang, X. & Zhan, S. Design of an ultra-broadband Films with a Memetic Algorithm and Mixed Integer Programming. ACS
near-perfect bilayer grating metamaterial absorber based on genetic Photonics 5, 684–691 (2018).
algorithm. Opt. Express 28, 15347–15359 (2020). 44. Wang, H., Zheng, Z., Ji, C. & Jay Guo, L. Automated multi-layer optical
18. Sonmez, F. O. & Tan, C. Structural optimization using simulated design via deep reinforcement learning. Mach. Learn.: Sci. Technol. 2,
annealing. Simulated annealing 2008, 281–306 (2008). 025013 (2021).
19. Sonmez, F. O. Shape optimization of 2D structures using simulated 45. Kitai, K. et al. Designing metamaterials with quantum annealing and
annealing. Computer methods Appl. Mech. Eng. 196, 3279–3299 factorization machines. Phys. Rev. Res. 2, 013319 (2020).
(2007). 46. Zhang, G. Quantum-inspired evolutionary algorithms: a survey and
20. Katoch, S., Chauhan, S. S. & Kumar, V. A review on genetic algorithm: empirical study. J. Heuristics 17, 303–351 (2011).
past, present, and future. Multimed. tools Appl. 80, 8091–8126 (2021). 47. Choe, I.-H. et al. Can quantum genetic algorithm really improve
21. Kadowaki, T. & Nishimori, H. Quantum annealing in the transverse quantum backpropagation neural network? Quantum Inf. Process. 22,
Ising model. Phys. Rev. E 58, 5355–5363 (1998). 154 (2023).
22. Tanaka, S., Tamura, R. & Chakrabarti, B. K. Quantum spin glasses, 48. Xu, Z. & Zhou, X. Determination of the Critical Slip Surface of Slope
annealing and computation. (Cambridge University Press, 2017). Based on the Improved Quantum Genetic Algorithm and Random
23. Albash, T. & Lidar, D. A. Adiabatic quantum computation. Rev. Mod. Forest. KSCE J. Civ. Eng. 26, 2126–2138 (2022).
Phys. 90, 015002 (2018). 49. Shu, W. & He, B. 169–176 (Springer).

npj Computational Materials | (2024)10:257 10

[Link] Article

50. Saad, H. M. H., Chakrabortty, R. K., Elsayed, S. & Ryan, M. J. resources of the Oak Ridge Leadership Computing Facility at the Oak Ridge
Quantum-Inspired Genetic Algorithm for Resource-Constrained National Laboratory, which is supported by the Office of Science of the U.S.
Project-Scheduling. IEEE Access 9, 38488–38502 (2021). Department of Energy under Contract No. DE-AC05-00OR22725. The
51. Kim, S., Jung, S., Bobbitt, A., Lee, E. & Luo, T. Wide-angle spectral authors also would like to thank the Notre Dame Center for Research
filter for energy-saving windows designed by quantum annealing- Computing for supporting all the simulations in this work. Notice: This
enhanced active learning. Cell Rep. Phys. Sci. 5, 101847 (2024). manuscript has in part been authored by UT-Battelle, LLC under Contract
52. Jalas, D. et al. What is—and what is not—an optical isolator. Nat. No. DE-AC05-00OR22725 with the U.S. Department of Energy. The United
Photonics 7, 579–582 (2013). States Government retains and the publisher, by accepting the article for
53. Zhang, T. & Luo, T. Giant thermal rectification from polyethylene publication, acknowledges that the U.S. Government retains a non-exclu-
nanofiber thermal diodes. Small 11, 4657–4665 (2015). sive, paid up, irrevocable, world-wide license to publish or reproduce the
54. Shrestha, R. et al. Dual-mode solid-state thermal rectification. Nat. published form of the manuscript, or allow others to do so, for U.S. Gov-
Commun. 11, 4346 (2020). ernment 15 purposes. The Department of Energy will provide public access
55. Kim, S. et al. Quantum annealing-aided design of an ultrathin- to these results of federally sponsored research in accordance with the DOE
metamaterial optical diode. Nano Convergence 11, 16 (2024). Public Access Plan ([Link]
56. Moharam, M., Gaylord, T., Grann, E. B. & Pommet, D. A. Formulation for
stable and efficient implementation of the rigorous coupled-wave Author contributions
analysis of binary gratings. J. Optical Soc. Am. A 12, 1068–1076 Z.X., E.L., and T.L. conceived the idea; Z.X. designed and performed the
(1995). QGA code for the research; W.S. and S.K. performed the QA for the research;
57. Preskill, J. Quantum computing in the NISQ era and beyond. Quantum A.B. performed the energy saving calculations; Z.X. wrote most parts of the
2, 79 (2018). manuscript; All authors discussed the results and commented on the
58. Shor, P. W. Quantum computing. Doc. Mathematica 1, 1 (1998). manuscript.
59. Yang, S. Y., Liu, F. & Jiao, L. C. Novel genetic algorithm based on the
quantum chromosome. J. Xidian Univ. 31, 76–81 (2004). Competing interests
60. Nielsen, M. A. & Chuang, I. L. Quantum computation and quantum The authors declare no competing interests.
information. (Cambridge University Press, 2010).
61. DiVincenzo, D. P. The Physical Implementation of Quantum Additional information
Computation. Fortschr. der Phys. 48, 771–783 (2000). Supplementary information The online version contains
62. Schuld, M., Sinayskiy, I. & Petruccione, F. The quest for a Quantum supplementary material available at
Neural Network. Quantum Inf. Process. 13, 2567–2586 (2014). [Link]
63. Zhang, G. & Rong, H. in Computational Science – ICCS 2007. (eds Y.
Shi, G. D. van Albada, J. Dongarra, & P. M. A. Sloot) 243-250 (Springer Correspondence and requests for materials should be addressed to
Berlin Heidelberg). Eungkyu Lee or Tengfei Luo.
64. Lewis, R. J. (Citeseer).
65. Steinberg, D. & Colla, P. CART: classification and regression trees. Reprints and permissions information is available at
top. ten algorithms data Min. 9, 179 (2009). [Link]
66. Breiman, L. Random Forests. Mach. Learn. 45, 5–32 (2001).
67. Lagomarsino, D., Tofani, V., Segoni, S., Catani, F. & Casagli, N. A Tool Publisher’s note Springer Nature remains neutral with regard to
for Classification and Regression Using Random Forest jurisdictional claims in published maps and institutional affiliations.
Methodology: Applications to Landslide Susceptibility Mapping and
Soil Thickness Modeling. Environ. Modeling Assess. 22, 201–214 Open Access This article is licensed under a Creative Commons
(2017). Attribution-NonCommercial-NoDerivatives 4.0 International License,
68. Matin, S. S., Farahzadi, L., Makaremi, S., Chelgani, S. C. & Sattari, G. which permits any non-commercial use, sharing, distribution and
Variable selection and prediction of uniaxial compressive strength and reproduction in any medium or format, as long as you give appropriate
modulus of elasticity by random forest. Appl. Soft Comput. 70, credit to the original author(s) and the source, provide a link to the Creative
980–987 (2018). Commons licence, and indicate if you modified the licensed material. You
69. Mercadier, M. & Lardy, J.-P. Credit spread approximation and do not have permission under this licence to share adapted material
improvement using random forest regression. Eur. J. Operational Res. derived from this article or parts of it. The images or other third party
277, 351–365 (2019). material in this article are included in the article’s Creative Commons
70. Peng, C., Li, Y., Cao, L. & Jiao, L. in 2019 IEEE congress on licence, unless indicated otherwise in a credit line to the material. If material
evolutionary computation (CEC). 1060–1067 (IEEE). 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,
Acknowledgements you will need to obtain permission directly from the copyright holder. To
This research was supported by the Quantum Computing Based on view a copy of this licence, visit [Link]
Quantum Advantage Challenge Research (RS-2023-00255442) through the nc-nd/4.0/.
National Research Foundation of Korea (NRF) funded by the Korean
government (Ministry of Science and ICT(MSIT)). This research also used © The Author(s) 2024

npj Computational Materials | (2024)10:257 11