Article

Machine Learning-Assisted Prediction of Ambient-Processed

Perovskite Solar Cells’ Performances
Dowon Pyun 1,† , Seungtae Lee 1,† , Solhee Lee 1,† , Seok-Hyun Jeong 1 , Jae-Keun Hwang 1 , Kyunghwan Kim 1 ,
Youngmin Kim 2 , Jiyeon Nam 1 , Sujin Cho 1 , Ji-Seong Hwang 1 , Wonkyu Lee 1 , Sangwon Lee 2 , Hae-Seok Lee 2 ,
Donghwan Kim 1 and Yoonmook Kang 2, *

1 Department of Materials Science and Engineering, Korea University, Seoul 02841, Republic of Korea;
[email protected] (D.P.); [email protected] (S.L.); [email protected] (S.L.);
[email protected] (S.-H.J.); [email protected] (J.-K.H.); [email protected] (K.K.);
[email protected] (J.N.); [email protected] (S.C.); [email protected] (J.-S.H.);
[email protected] (W.L.); [email protected] (D.K.)
2 Graduate School of Energy and Environment (KU-KIST Green School), Korea University, Seoul 02841,
Republic of Korea; [email protected] (Y.K.); [email protected] (S.L.); [email protected] (H.-S.L.)
* Correspondence: [email protected]; Tel.: +82-2-3290-3713
† These authors contributed equally to this work.

Abstract: As we move towards the commercialization and upscaling of perovskite solar cells, it
is essential to fabricate them in ambient environment rather than in the conventional glove box
environment. The efficiency of ambient-processed perovskite solar cells lags behind those fabricated
in controlled environments, primarily owing to external environmental factors such as humidity
and temperature. In the case of device fabrication in ambient environments, relying solely on
a single parameter, such as temperature or humidity, is insufficient for accurately characterizing
environmental conditions. Therefore, the dew point is introduced as a parameter which accounts for
both temperature and humidity. In this study, a machine learning model was developed to predict
the efficiency of ambient-processed perovskite solar cells based on meteorological data, particularly
the dew point. A total of 238 perovskite solar cells were fabricated, and their photovoltaic parameters
Citation: Pyun, D.; Lee, S.; Lee, S.; and dew points were collected from March to December 2023. The collected data were used to train
Jeong, S.-H.; Hwang, J.-K.; Kim, K.; various tree-based machine learning models, with the random forest model achieving the highest
Kim, Y.; Nam, J.; Cho, S.; Hwang, J.-S.; accuracy. The efficiencies of the perovskite solar cells fabricated in January and February 2024 were
et al. Machine Learning-Assisted predicted with a MAPE of 4.44%. An additional Shapley Additive exPlanations analysis confirmed
Prediction of Ambient-Processed the significance of the dew point in the performance of perovskite solar cells.
Perovskite Solar Cells’ Performances.
Energies 2024, 17, 5998. https://
Keywords: ambient processed; perovskite solar cells; dew point; artificial intelligence; machine learning
doi.org/10.3390/en17235998

Academic Editors: Hubert Seigneur

and Kristopher Davis
1. Introduction
Received: 4 November 2024
Revised: 26 November 2024 Perovskite solar cells have attracted significant attention as a next-generation pho-
Accepted: 27 November 2024 tovoltaic technology due to their high efficiency [1], bandgap tunability [2], and high
Published: 28 November 2024 absorption coefficients [3]. Over the past decade, their efficiency has notably increased from
3.81% [4] in 2009 to 26.7% in 2024 [5], and have now almost approached the highest reported
efficiency of silicon solar cells. With the rapid growth in their efficiency, perovskite solar
cells have also been integrated into perovskite/silicon tandem solar cells [6–10], which are
Copyright: © 2024 by the authors.
expected to surpass the theoretical efficiency limits of single-junction solar cells. Traditional
Licensee MDPI, Basel, Switzerland.
solution-based methods such as spin-coating are generally utilized for perovskite forma-
This article is an open access article
tion [11,12]. The dry deposition of perovskite has also emerged as a promising method
distributed under the terms and
recently [13–15]. However, most of the high-efficiency perovskite solar cells reported have
conditions of the Creative Commons
been achieved at the lab scale with small active areas (less than 0.1 cm2 ) under controlled
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
environments [16] such as glove boxes filled with inert gas. This is attributed to the high
4.0/).
sensitivity of perovskite materials to external environmental conditions such as moisture.

Energies 2024, 17, 5998. https://doi.org/10.3390/en17235998 https://www.mdpi.com/journal/energies

Energies 2024, 17, 5998 2 of 12

However, for the future commercialization of large-scale perovskite solar cells, fab-
rication techniques must advance beyond conventional glove box processing to allow
for fabrication in an ambient environment. Thus, many research groups are working to
enhance the performance and stability of ambient-environment-processed perovskite solar
cells using various strategies [17–22]. Unfortunately, those fabricated in ambient environ-
ments generally exhibit a relatively low performance [23]. An effective method to mitigate
the influence of the external environment on the performance of perovskite solar cells to
achieve an efficiency of exceeding 25% under ambient atmosphere is still lacking. It is clear
that ambient environments contain various elements that can adversely affect perovskite
absorbers such as humidity, temperature, and light [24–27]. A high relative humidity, in
particular, disrupts the nucleation process during precursor spin-coating and annealing,
leading to negative impacts on the surface morphology of the perovskite film and the
device’s performance [23,28,29]. Moreover, it accelerates the irreversible decomposition
of perovskite into lead iodide [30]. Such degradation caused by the ambient environment
reduces the reproducibility of the fabricated devices, making it more difficult to predict
their photovoltaic performance.
In this work, we aim to construct a model to predict the efficiency of devices fabricated
under ambient environmental conditions. Recently, artificial intelligence (AI) has been ap-
plied to predict the efficiency of perovskite solar cells under varying external environmental
conditions using machine learning [31–33]. Most prediction models have only focused on
relative humidity among various external environmental factors under precisely controlled
conditions. However, in the case of device fabrication in ambient environments, we argue
that relying solely on a single parameter, such as temperature or humidity, is insufficient
for accurately characterizing environmental conditions. For instance, Ko et al. [34] reported
that different thin film qualities were obtained even at the same relative humidity. To
achieve a more accurate characterization of ambient environmental conditions, therefore,
the dew point was introduced as a parameter that accounts for both temperature and
humidity. The dew point is a parameter proportional to both relative humidity and temper-
ature, which will be discussed more in the following Discussion Section. To the best of our
knowledge, this is the first attempt to use dew points to predict the efficiency of perovskite
solar cells. Our prediction provides guidelines for fabricating ambient-processed perovskite
solar cells and is anticipated to contribute significantly to the future commercialization and
upscaling of perovskite solar cells

2. Materials and Methods

Fluorine-doped tin oxide (FTO) substrates were purchased from Pilkington. Acetone,
ethanol, and 2-propanol (99.7% purity; HPLC grade) were purchased from Duksan. N,N-
dimethylformamide (anhydrous, 99.8% purity), dimethyl sulfoxide (anhydrous, 99.9%
purity), diethyl ether (anhydrous, 99.7% purity), chlorobenzene (anhydrous, 99.8% purity),
Acetonitrile (anhydrous, 99.8% purity), bis (trifluoromethane) sulfonimide lithium salt
(Li-TFSI, 99.95% purity, trace metal basis), and 4-tert-butylpyridine (98% purity) were
purchased from Sigma-Aldrich. Methylammonium iodide (99.99% purity) was purchased
from Greatcell Solar Materials. 2,2′ ,7,7′ -tetrakis-(N,N-di-4-methoxyphenylamino)-9,9′ -
spiro bifluorene (spiro-OMeTAD, 99% purity) was purchased from Lumtec. Lead iodide
(PbI2 , 99.99% purity, trace metal basis, for perovskite precursor) was acquired from Tokyo
Chemical Industry (TCI, Chuo-ku, Japan).
FTO glass substrates were cleaned using acetone, ethanol, and 2-propanol sequentially
in an ultrasonicator for 15 min. The cleaned FTO substrate was then exposed to UV–ozone
for 30 min with UV-ozone cleaner (AH1700 model, AHTECH LTS Co., Ltd., Anyang-si,
Gyeonggi-do, Republic of Korea). SnO2 was deposited by RF magnetron sputtering, as
optimized in our previous study [11]. A MAPbI3 precursor solution was prepared by
dissolving PbI2 (1 mM), methyl ammonium iodide (172 mg), and dimethyl sulfoxide
(71 µL) in a dimethylformamide solvent (636.6 µL). Diethyl ether was dropped during spin
coating, and the substrate was dried on a hot plate at 60 ◦ C for 1 min, followed by 100 ◦ C
 Energies 2024, 17, x FOR PEER REVIEW 3 of 12
Energies 2024, 17, 5998 3 of 12

the5substrate
for min. The was dried on aofhot
fabrication theplate at 60 °Cabsorber
perovskite for 1 min,layer
followed by 100 to
is identical °C that
for 5reported
min. The
fabrication of the perovskite absorber layer is identical to that reported
in our previous research [10,35,36]. For the hole transportation layer (HTL) coating, the in our previous re-
search [10,35,36]. For the hole transportation layer (HTL) coating,
spiro-OMeTAD powder was dissolved in 4-tert-butylpyridine (28.8 µL), Li-TFSI (520 mg the spiro-OMeTAD powder
was
in 1 mLdissolved in 4-tert-butylpyridine
of acetonitrile), and chlorobenzene (28.8 µL),
(1 mL)Li-TFSI (520 mg
solutions. in 1solution
This mL of acetonitrile),
was appliedand to
chlorobenzene
the substrate, and (1 mL) solutions.
a thin layer with This200 solution was applied
nm thickness to the substrate,
was formed and a thin
by spin coating. layer
Finally,
with
100 nm 200ofnm
Au thickness was formed
was deposited by thermalby spin coating. Finally, 100 nm of Au was deposited by
evaporation.
thermal evaporation.
The photovoltaic performance of the perovskite solar cells was measured using a solar
The photovoltaic
simulator (WACOM WXS-155S10 performanceclass of theAAA)
perovskite solarmW/cm
with 100 cells was 2 irradiation
measured using usinga asolar
Xe
simulator
lamp. (WACOM
Modulated WXS-155S10
reference Si solarclass
cellsAAA) with 100
were used mW/cm irradiation
for calibration 2
before theusing a Xe lamp.
measurements.
Modulated
The devicesreference Si solar using
were evaluated cells were
shadowusedmasks
for calibration
with anbefore
aperturethe size
measurements.
of 0.075 cmThe 2 . The
de-
vicesvoltage
scan were evaluated
setting timeusingwas
shadow
40 mV/s.masks with an aperture size of 0.075 cm . The scan voltage
2

setting time was 40 mV/s.

3. Results and Discussion
3. Results and Discussion
Significant variations in precipitation levels exist regionally and seasonally. For in-
stance, South Korea
Significant experiences
variations concentrated
in precipitation rainfall
levels in July andand
exist regionally August, whereas
seasonally. ForPortugal
instance,
receives less rainfall during the same period. As precipitation is directly
South Korea experiences concentrated rainfall in July and August, whereas Portugal receives related to humidity,
the
lessefficiency of ambient-environment-processed
rainfall during the same period. As precipitation perovskite solar
is directly cells varies
related depending
to humidity, on
the effi-
the location and timing of their fabrication. Experimentally, perovskite
ciency of ambient-environment-processed perovskite solar cells varies depending on the loca- films fabricated in
an ambient environment under high-humidity conditions (>40%) exhibited
tion and timing of their fabrication. Experimentally, perovskite films fabricated in an ambient surface stains
(Figure S1) despite
environment underan air conditioner
high-humidity and dehumidifier
conditions (>40%) exhibitedbeingsurface
in operation. These S1)
stains (Figure stainsde-
are the result of the non-uniform crystallization of a perovskite precursor.
spite an air conditioner and dehumidifier being in operation. These stains are the result of the Since the per-
ovskite precursor
non-uniform and final of
crystallization thin films are exposed
a perovskite precursor. toSince
the external environment
the perovskite precursor during the
and final
spin-coating and annealing steps, it is clear that external environmental
thin films are exposed to the external environment during the spin-coating and annealing factors significantly
influence
steps, it isnot
clearonlythatthe crystallization
external environmentalstep but also the
factors device’s properties.
significantly influence not only the crys-
In this study, perovskite solar cells were fabricated, and their photovoltaic parameters
tallization step but also the device’s properties.
and meteorological data at the time of fabrication were collected from March to December
In this study, perovskite solar cells were fabricated, and their photovoltaic parameters
2023. The collected data were used to train a machine learning model to predict the device’s
and meteorological data at the time of fabrication were collected from March to December
efficiency at specific dew points. The dew point for a given relative humidity (RH) and air
2023. The collected data were used to train a machine learning model to predict the device’s
temperature (T) can be calculated using the Magnus formula [34].
efficiency at specific dew points. The dew point for a given relative humidity (RH) and air
temperature (T) can be calculated using the Magnus formula  [34].
RH 17.62 × T

243.12 × ln 100 + 243.12 +T
Dew Point (◦ C) = f(RH, T) = .   .
. × T

Dew Point °C f RH, T17.62 − ln 100 + 243.12
RH 17.62
. +T
.
.

The formula shows

The showsthat
thatthe
thedew
dewpoint is aisparameter
point a parameterproportional to relative
proportional humidity
to relative hu-
and temperature.
midity Figure Figure
and temperature. 1 shows the flowchart
1 shows for the for
the flowchart dew-point-based efficiency
the dew-point-based predic-
efficiency
tion process.
prediction The detailed
process. processes
The detailed are discussed
processes in thein
are discussed following sections.
the following sections.

Figure1.1.Entire
Figure Entireflowchart
flowchartfor
fordew-point-based
dew-point-basedefficiency
efficiencyprediction
predictionininthis
thisstudy.
study.
 Energies2024,
Energies 17,x5998
2024,17, FOR PEER REVIEW 4 of
4 of1212

3.1.Data
3.1. DataCollection
Collection
InInthis
thiswork,
work,totoinvestigate
investigatethe theeffect
effectofofexternal
externalenvironments
environmentson onthe theperovskite
perovskite
layer, perovskite solar cells were fabricated with structure of FTO/SnO
layer, perovskite solar cells were fabricated with structure of FTO/SnO2/MAPbI33/spiro- 2 /MAPbI /spiro-
OMeTAD/Au. Their photovoltaic parameters and meteorological data at the time of fabrica-of
OMeTAD/Au. Their photovoltaic parameters and meteorological data at the time
fabrication
tion were collected
were collected over a year over a year (March–December
(March–December 2023). These 2023). Theseused
data were datatowere trainused
a ma-to
chine learning model to predict the device’s efficiency at specific dew points. Dew pointspoints.
train a machine learning model to predict the device’s efficiency at specific dew were
Dew points
collected usingwere
two collected
methods: using two methods:
(1) laboratory (1) laboratory
measurements and (2)measurements
data from the Korea and (2) data
Mete-
from the Korea Meteorological Administration (KMA). Relative humidity
orological Administration (KMA). Relative humidity and temperature during fabrication were and temperature
during fabrication were recorded using a thermohygrometer in our laboratory (Figure S2).
recorded using a thermohygrometer in our laboratory (Figure S2). Dew points were calculated
Dew points were calculated using the Magnus equation (Equation (1)). Dew points were
using the Magnus equation (Equation (1)). Dew points were also obtained from the KMA
also obtained from the KMA (www.weather.go.kr, accessed on 29 December 2023). A
(www.weather.go.kr, accessed on 29 December 2023). A linear relationship was confirmed be-
linear relationship was confirmed between the laboratory and KMA dew points (Figure S3).
tween the laboratory and KMA dew points (Figure S3). In both cases, higher dew points were
In both cases, higher dew points were observed during summer, and lower dew points
observed during summer, and lower dew points were observed during winter, validating the
were observed during winter, validating the feasibility of using KMA-reported dew points.
feasibility of using KMA-reported dew points. Given its broader applicability, the KMA-re-
Given its broader applicability, the KMA-reported dew points were used as input data for
ported dew points were used as input data for the machine learning.
the machine learning.
Figure 2 shows the photovoltaic parameters plotted against dew points. Additionally,
Figure 2 shows the photovoltaic parameters plotted against dew points. Additionally,
only the external environment was varied during the perovskite film fabrication, while all
only the external environment was varied during the perovskite film fabrication, while all
other steps, including measurements, were conducted under identical standard test condi-
other steps, including measurements, were conducted under identical standard test conditions.
tions.
DeviceDevice performance
performance tended tended to decrease
to decrease with increasing
with increasing dew consistent
dew points, points, consistent with
with previous
previous research indicating high humidity’s adverse effects on perovskite
research indicating high humidity’s adverse effects on perovskite solar cells. One important solar cells. One
important observation is that the decline in the open-circuit voltage (V OC) and fill factor (FF)
observation is that the decline in the open-circuit voltage (VOC ) and fill factor (FF) was
was
moremore significant
significant than thanthethe decline
decline in the
in the short-circuit
short-circuit current
current density
density (JSC
(JSC ). ).This
Thisbehavior
behavioris
isconsistent
consistentwith
withthe
the effect
effect of
of defect
defect density
density on on the
the photovoltaic
photovoltaicperformances
performancesofofthin-filmthin-film
solar cells. Related findings have been reported in various
solar cells. Related findings have been reported in various studies, studies, including those on
including CdTe
those on
solar
CdTe solar cells [37–40] and perovskite solar cells [41–43]. For instance, Chouhan et al. re-
cells [37–40] and perovskite solar cells [41–43]. For instance, Chouhan et al. [43] [43]
ported
reportedsimulation
simulationresults
resultsdemonstrating
demonstrating that increases
that increases inin
interface
interface and
andbulk
bulkdefect
defectdensity
density
significantly
significantlyimpact
impactVV OC and
OC and FF. In
FF. Infact,
fact,many
many groups
groups employ
employ interface
interface [11]
[11] and
and bulk
bulk [44]
[44]
passivation
passivationtechnologies
technologiestotomitigate
mitigatedrops
dropsin inthe
theperformance
performance of of perovskite
perovskite absorbers.
absorbers.

Photovoltaicparameters
Figure2.2.Photovoltaic
Figure parametersofofperovskite
perovskitesolar
solarcells
cellsfabricated
fabricatedunder
undervarious
variousdew
dewpoints.
points.AA
total of 238 devices were collected and measured. (a) V , (b) J , (c) FF, and (d) efficiency.
total of 238 devices were collected and measured. (a) VOC , (b) JSC , (c) FF, and (d) efficiency.
OC SC
 Energies 2024, 17, 5998 5 of1212
Energies 2024, 17, x FOR PEER REVIEW 5 of

Additional analyses were conducted using the Solar Cell Capacitance Simulator
Additional analyses were conducted using the Solar Cell Capacitance Simulator (SCAPS-
(SCAPS-1D) simulation program [45]. The bulk defect density in MAPbI3 was controlled
1D) simulation program [45]. The bulk defect density in MAPbI3 was controlled to investigate
to investigate how it affects device performance. The structure of the device used in the
how it affectsisdevice
simulation performance.
identical to that of The structure ofdevice,
the fabricated the device used in the
specifically, simulation is iden- ,
FTO/SnO2/MAPbI 3
tical to that of the fabricated device, specifically, FTO/SnO2/MAPbI
spiro-OMeTAD/Au. The input parameters are summarized in Table S1. Those parameters 3, spiro-OMeTAD/Au. The

input parameters
are based on those arereported
summarized in Table[46]
in reference S1. Those
and theparameters
experimentalare based
valuesonconfirmed
those reported
in our
inprevious
reference [46] and the experimental values confirmed in our previous research,
research, such as a bandgap of 1.52 eV for MAPbI3 [10]. As a result, an efficiency such as a
bandgap of 1.52 eV for MAPbI 3 [10]. As a result, an efficiency
15 − of
3
of about 20% was obtained at a density of 5 × 10 cm . Considering that a densityabout 20% was obtained at a
density of 5 × 10 15
15 cm .
−3− Considering
3 that a density level of 8 × 10 15 cm−3 was confirmed in our
level of 8 × 10 cm was confirmed in our previous study [11], it was confirmed that
previous study [11],
the simulation it was
results confirmed
align well with thatthe
theexperimental
simulation results
data. align well with
As shown the experi-
in Figure 3, an
mental data. As shown in Figure 3, an increasing defect density led to
increasing defect density led to a deterioration in the performance of perovskite solar a deterioration incells,
the
performance
particularly of inperovskite
VOC (Figure solar
3b)cells,
and particularly
FF (Figure 3d).in VOC (Figure 3b) and FF (Figure 3d).

Figure
Figure3.3.Effects
Effectsofofbulk
bulkdefect density
defect in in
density MAPbI
MAPbI 3 simulated
3 simulatedwith SCAPS-1D.
with SCAPS-1D. (a) J-V curve,
(a) J-V (b) V
curve, (b) , (c)
OCV OC ,
JSC, (d)
(c) JSCFF, and
, (d) FF,(e)and
efficiency. The redThe
(e) efficiency. arrowredinarrow
(a) indicates an increaseanofincrease
in (a) indicates bulk defect density
of bulk in MAPbI
defect 3.
density
in MAPbI3 .
Since defects located in the perovskite layer act as a recombination site, this defect-me-
diatedSince defects located
recombination in reduced
leads to the perovskite
carrier layer act asTherefore,
collection. a recombination site, this
deteriorated defect-
photovol-
mediated
taic recombination
parameters leadsas
were observed, tothe
reduced
numbercarrier collection.
of defects Therefore,
increases. deteriorated
This trend, similar tophoto-
that
voltaic parameters
observed in Figure 2,were observed,
indicates that as
thethe
dewnumber of defects
point affects theincreases.
perovskiteThis trend,during
absorber similar
fabrication, suggesting its potential as a predictor of perovskite solar cell performance. In
Energies 2024, 17, 5998 6 of 12

to that observed in Figure 2, indicates that the dew point affects the perovskite absorber
during fabrication, suggesting its potential as a predictor of perovskite solar cell perfor-
mance. In fact, initially, at the low dew point, the data exhibit relatively similar trends;
however, at higher dew points, the distribution among data points increased, deviating
from the simulation results. This inconsistency between the experiment data and the simu-
lation results may be attributed to the low stability of devices fabricated under high dew
point conditions. Devices fabricated under high-dew-point conditions exhibited signifi-
cantly deteriorated photovoltaic performances with low stabilities, resulting in considerable
distribution between data points even fabricated under the same dew point.
Combining the experimental data and simulation results, we made the following
conclusions. Perovskite films crystallized in high-dew-point conditions tend to incorporate
numerous defects. These internal bulk defects lead to degraded photovoltaic performances.
As a result, perovskite films fabricated under high-dew-point conditions result in deterio-
rated photovoltaic performances.

3.2. Machine Learning

Due to the high sensitivity of perovskite material to various external factors during
ambient atmosphere fabrication, efficiency deviations occurred even within the same batch.
Thus, data pre-processing was required for more accurate predictions. Outliers were
excluded from the experimental data, retaining only data within a 1.5% range of the batch’s
average efficiency value. This process resulted in 165 data points for constructing the
training and testing datasets.
The dataset was divided into training and test subsets in an 8:2 ratio, employing
10-fold cross-validation to train the models. This method partitions the dataset into ten
equally sized subsets, where nine subsets are used for training, and the remaining subset
serves as the test set. The process is repeated 10 times, with each of the ten subsets being
used exactly once as the test set. This approach enhances model learning efficiency, reduces
overfitting risk, and improves validation confidence, especially with limited data.
Model performance was validated using the test dataset, with predictive capabilities
assessed through the coefficient of determination (R2 ), root mean square error (RMSE),
and mean absolute percentage error (MAPE). These metrics are crucial for evaluating
model precision and reliability. Table 1 lists the calculation methods used for each perfor-
mance metric.

Table 1. Performance metrics with equations for evaluating machine learning predictions. For more
detail, n is the total number of observations, yi is the actual observed value for the ith observation, ŷi
is the predicted value for the ith observation by the machine learning model, and y is the mean of the
actual observed values.

Performance Metrics Equation

n
R2 ∑i=1 (yi −ŷi )
1− n
∑ i =1 ( y i − y )
s
n
RMSE 1 2
n ∑ (yi − ŷi )
i =1

MAPE 1 n |ŷi − yi |
∑ × 100
n i =1 yi

The model selection process emphasized tree-based models, based on Grinsztajn

et al. [47], indicating that tree-based ensemble methods, such as bagging and boosting,
are more effective than deep learning models on tabular data and that deep learning
struggles to detect patterns within small datasets. Consequently, five distinct machine-
learning algorithms were adopted: the decision tree regressor (DT), random forest regressor
(RF), gradient boosting regressor (GB), light gradient boosting machine (LightGBM), and
adaptive boosting regressor (AdaBoost). The best model was selected through performance
 Energies 2024, 17, 5998 7 of 12
FOR PEER REVIEW 7 of 12

comparison after hyperparameter optimization. Table 2 lists the prediction errors for each
model’s efficiency.
random forest was selected as the prediction model. Figure 4a shows predicted versus
Table 2. Prediction accuracy with each machine learning model.
actual efficiency values for the training and test sets, confirming the model’s accuracy and
ability to avoid overfitting (see Figure S4 for remaining
Model R2 photovoltaic
RMSE [%] parameters).
MAPE [%]
RF 0.9349 0.9434 5.14
Table 2. Prediction GB
accuracy with each machine 0.9314
learning model. 0.9683 5.25
DT 0.9292 0.9638 5.29
AdaBoost 0.8628 1.3690 7.6
Model R2 RMSE [%] MAPE [%]
LightGMB 0.8160 1.5856 9.61
RF 0.9349 0.9434 5.14
GB The random0.9314
forest model outperformed0.9683others, with the highest5.25
R2 value of 0.9349
DT and the lowest0.9292
RMSE and MAPE values0.9638 5.29
of 0.9434% and 5.14%, respectively. Therefore,
AdaBoost random forest0.8628
was selected as the prediction
1.3690model. Figure 4a shows predicted
7.6 versus
actual efficiency values for the training and test sets, confirming the model’s accuracy and
LightGMB 0.8160 1.5856 9.61
ability to avoid overfitting (see Figure S4 for remaining photovoltaic parameters).

Figure 4. PredictionFigure
results
4. using theresults
Prediction trained random
using forest
the trained model.
random (a)model.
forest Graph(a)showing the actual
Graph showing the actual
values on the x-axis and the predicted values on the y-axis; closer alignment to
values on the x-axis and the predicted values on the y-axis; closer alignment to the y = x line indicatesthe y = x line indicates
more accurate predictions. The light red dots represent training dataset predictions, while dark dark
more accurate predictions. The light red dots represent training dataset predictions, while red red
dots represent test dataset predictions. (b) The efficiency distribution (box
dots represent test dataset predictions. (b) The efficiency distribution (box chart) of the fabricated chart) of the fabricated
devices in January and February 2024, with predicted results (red stars) obtained using our model.
devices in January and February 2024, with predicted results (red stars) obtained using our model.
The performance of perovskite solar cells fabricated in January and February 2024
The performance of perovskite
was compared solarpredictions
with model cells fabricated inthe
to validate January and experiments
model. Five February 2024 over two
months showed that the predicted efficiencies (red stars
was compared with model predictions to validate the model. Five experiments over two in Figure 4b) fell within the actual
efficiency range, achieving an MAPE of 4.44%.
months showed that the predicted efficiencies (red stars in Figure 4b) fell within the actual
This outcome is significant considering Korea’s climate conditions, with high dew
efficiency range, achieving
points in Julyan MAPE
and Augustof 4.44%.
and low dew points in January and February. The model,
This outcome trained
is significant
on dataconsidering
from March to Korea’s
December climate conditions,
(periods with high dew
typically characterized pointsdew
by higher
in July and August and low dew points in January and February. The model, trained on data
points), effectively predicted solar cell parameters even during low dew point seasons,
demonstrating reliable predictions across diverse environmental
from March to December (periods typically characterized by higher dew points), effectively conditions; this highlights
its potential as a valuable tool for guiding the fabrication of perovskite solar cells in real-
predicted solar cellworld
parameters
settings.
even during low dew point seasons, demonstrating reliable
predictions across diverse Theenvironmental conditions;
complexity of tree-based ensemblethis models
highlights its potential
necessitates as amethodologies
additional valuable
tool for guiding theforfabrication of perovskite
interpretability. solar cells
Shapley Additive in real-world
exPlanations (SHAPs)settings.
are one such tool, facilitating
model interpretation [48–50]. This study used SHAPs to
The complexity of tree-based ensemble models necessitates additional methodologiesidentify key factors influencing
for
perovskite solar cell efficiency. The graph on the left in Figure 5a shows feature importance,
interpretability. Shapley Additive exPlanations (SHAPs) are one such tool, facilitating model
interpretation [48–50]. This study used SHAPs to identify key factors influencing perovskite
solar cell efficiency. The graph on the left in Figure 5a shows feature importance, revealing
the impact of various factors on the power conversion efficiency (PCE). The feature im-
portance (average SHAP value) of the dew point is 2.63, higher than that of the other values,
 Energies 2024, 17, 5998 8 of 12

revealing the impact of various factors on the power conversion efficiency (PCE). The
feature importance (average SHAP value) of the dew point is 2.63, higher than that of
the other values, indicating that the dew point is more important than other factors in
predicting the PCE; this suggests that, in real-world production environments, the influence
Energies 2024, 17, x FOR PEER REVIEWof dew point may outweigh other factors. The graph on the right in Figure 5a shows a violin 8 of 12
plot, visually representing the SHAPs distribution for each sample. The plot thickness
reflects the concentration of points while utilizing a color gradient, where red indicates high
feature values and blue represents low feature values. The graph shows a clear tendency
increases; thisofindicates
for the color the violinthat anrepresenting
plot, increase in the
the dew
dew point
point,negatively impacts
to shift from red to efficiency.
blue as The
the SHAPs value increases; this indicates that an increase in the dew point
negative relationship between dew point and efficiency is even more evident in Figure 5b, negatively
impacts
which efficiency.
suggests thatThe negativethe
performing relationship
process atbetween dew below
a dew point point and efficiency
13 °C is even
is beneficial for achiev-
more evident in Figure 5b, which suggests that performing the process at a dew point below
ing ◦high-efficiency ambient-processed perovskite solar cells. Furthermore, as shown in Figure
13 C is beneficial for achieving high-efficiency ambient-processed perovskite solar cells.
S5, the threshold derived from SHAP analysis aligns with the threshold obtained from model
Furthermore, as shown in Figure S5, the threshold derived from SHAP analysis aligns with
predictions,
the thresholdvalidating the consistency
obtained from of these
model predictions, findings.the consistency of these findings.
validating

Figure 5.
Figure (a)Feature
5. (a) Feature importance
importance in inthe
themodel.
model.Dew Dewpoint
pointshows
showsthe thehighest
highest feature
featureimportance,
importance, sug-
suggesting that dew point is the crucial factor with a substantial impact on efficiency. (b) Relationship
gesting that dew point is the crucial factor with a substantial impact on efficiency. (b) Relationship
between SHAPs
between SHAPs values
values and
anddewdewpoints.
points.Blue
Bluebackground
background region
region implies
impliesthethe
trend of the
trend obtained
of the obtained data
data points. The point where SHAP is zero is indicated with a red line, and
points. The point where SHAP is zero is indicated with a red line, and the dashed linethe dashed line indicates
indicates the
the criteria for dew point suggested in this work that exhibits a positive
criteria for dew point suggested in this work that exhibits a positive SHAP value. SHAP value.

4. Conclusions
In this study, we developed a machine learning model to predict the efficiency of am-
bient-processed perovskite solar cells based on the dew point. A total of 238 perovskite solar
cells were fabricated, and their photovoltaic parameters were collected along with the dew
Energies 2024, 17, 5998 9 of 12

4. Conclusions
In this study, we developed a machine learning model to predict the efficiency of
ambient-processed perovskite solar cells based on the dew point. A total of 238 perovskite
solar cells were fabricated, and their photovoltaic parameters were collected along with the
dew points at the time of fabrication. We observed a correlation where devices fabricated in
high-dew-point environments exhibited lower efficiency. Using the SCAPS-1D simulation
program, we indirectly confirmed that the dew point influenced the perovskite layer
during fabrication.
Various tree-based models (DT, ET, RF, GB, and AdaBoost) were applied for compara-
tive analysis. The random forest model demonstrated the highest prediction accuracy for
efficiency. Using the model trained with data from March to December 2023, efficiency
predictions for January and February 2024 were made, achieving a MPAE of 4.44%; this
underscores the model’s capability to accurately predict the efficiency of perovskite solar
cell fabricated in an ambient atmosphere with a dew point.
Certainly, there are regional and seasonal differences among laboratories, as well as
variations in perovskite fabrication methods such as whether additives are used. Therefore,
it could be challenging to expect the same level of prediction accuracy if the data points
obtained in our laboratory were applied in other laboratories. In addition, considering
the numerous variables present in the ambient environment, there may be other potential
environmental factors that are not considered in this study. Although every effort was made
to ensure that the experiment processes (substrate cleaning, precursor solution synthesis,
and measurements) were performed under strictly identical conditions, factors such as total
duration of exposure to the ambient environment until device fabrication is completed,
the shelf life of materials, and barometric pressure which varies with altitude may also
influence the device performance. Nevertheless, the methodology proposed in this study
using the dew point holds significant potential for predicting the efficiency of ambient-
processed perovskite solar cells. It becomes possible to reliably predict the efficiency of
perovskite solar cells fabricated in ambient conditions, which could serve as a guideline for
future cell production under similar conditions. Moreover, the methodology applied in this
work offers the following additional directions for future development:
(1) From an industry perspective, the model could assist in optimizing production output.
Variations in perovskite solar cell efficiency under ambient conditions are closely
linked to production reliability, which, in turn, can significantly impact overall pro-
duction yields. Since the issue of the stability of perovskite solar cells being influenced
by environmental factors remains unresolved, employing this model to predict effi-
ciencies in advance and adjust production output represents a practical and effective
solution. Moreover, this capability is anticipated to play a pivotal role in enabling
the implementation of smart factory systems, which are defined by autonomous and
adaptive production processes.
(2) The methodology can be applied to various compositions of perovskite solar cells, not
only for MAPbI3 . Appling the same methodology to various compositions is expected
to provide more reliable and comprehensive results. Furthermore, this approach has
the potential to go beyond predicting cell efficiency, extending to forecasts of module-
level power output and establishing a framework for comprehensive predictive tools
applicable across the industry.
(3) If this methodology were applied in reverse, it could aid in developing perovskite solar
cells that are more resistant to environmental factors. Once this model is expanded to
accommodate various types of perovskite solar cells, it could be utilized to identify
cell types that exhibit the least sensitivity to changes in environmental conditions
by varying environmental parameters and analyzing the corresponding machine
learning predictions. This indicates that the model could also play a pivotal role in
proposing optimized compositions, materials, and structural designs for perovskite
solar cells. Such applications would further contribute to advancing research on the
stability of perovskite solar cells.
Energies 2024, 17, 5998 10 of 12

Considering all of these points, the model and methodology proposed in this work
are expected to make a significant contribution to the advancement of commercialization
and upscaling of perovskite solar cells in the future.

Supplementary Materials: The following supporting information can be downloaded at: https:
//www.mdpi.com/article/10.3390/en17235998/s1, Figure S1: Perovskite film fabricated under
different relative humidity conditions, Figure S2: Obtaining the meteorological data of the laboratory
using a thermohygrometer, Figure S3: Correlation between two different dew points, Figure S4.
Prediction results using the randon forest model, Figure S5. Validation of the SHAP-derived dew
point threshold, Table S1: Input parameters used in SCAPS-1D simulation.
Author Contributions: D.P., S.L. (Seungtae Lee) and S.L. (Solhee Lee) contributed equally to this
work. Conceptualization, D.P., S.L. (Seungtae Lee) and S.L. (Solhee Lee); methodology, D.P. and S.L.
(Solhee Lee); validation, S.L. (Solhee Lee); formal analysis, S.-H.J. and S.L. (Solhee Lee); investigation,
D.P. and J.-K.H.; resources, S.L. (Sangwon Lee), J.-K.H., J.N. and K.K.; data curation, K.K., Y.K.
(Youngmin Kim), S.C., J.-S.H. and W.L.; writing—original draft preparation, D.P. and S.L. (Seungtae
Lee); writing—review and editing, D.P. and S.L. (Seungtae Lee); visualization, D.P. and S.L. (Seungtae
Lee); supervision, D.K. and Y.K. (Yoonmook Kang); project administration, D.K. and Y.K. (Yoonmook
Kang); funding acquisition, H.-S.L. and D.K. All authors have read and agreed to the published
version of the manuscript.
Funding: This work was supported by the “Human Resources Program in Energy Technology” of the
Korea Institute of Energy Technology Evaluation and Planning (KETEP), granted financial resources
from the Ministry of Trade, Industry &Energy, Republic of Korea (No. 20204010600470). This work
was supported by the Korea Institute of Energy Technology Evaluation and Planning (KETEP) from
the Ministry of Trade, Industry, and Energy (20214000000680).
Data Availability Statement: The original contributions presented in the study are included in the
article, further inquiries can be directed to the corresponding author.
Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Min, H.; Lee, D.Y.; Kim, J.; Kim, G.; Lee, K.S.; Kim, J.; Paik, M.J.; Kim, Y.K.; Kim, K.S.; Kim, M.G. Perovskite solar cells with
atomically coherent interlayers on SnO2 electrodes. Nature 2021, 598, 444–450. [CrossRef] [PubMed]
2. Noh, J.H.; Im, S.H.; Heo, J.H.; Mandal, T.N.; Seok, S.I. Chemical management for colorful, efficient, and stable inorganic–organic
hybrid nanostructured solar cells. Nano Lett. 2013, 13, 1764–1769. [CrossRef] [PubMed]
3. Yin, W.-J.; Yang, J.-H.; Kang, J.; Yan, Y.; Wei, S.-H. Halide perovskite materials for solar cells: A theoretical review. J. Mater. Chem.
A 2015, 3, 8926–8942. [CrossRef]
4. Kojima, A.; Teshima, K.; Shirai, Y.; Miyasaka, T. Organometal halide perovskites as visible-light sensitizers for photovoltaic cells.
J. Am. Chem. Soc. 2009, 131, 6050–6051. [CrossRef] [PubMed]
5. NREL. Best Research-Cell Efficiencies. Available online: https://www.nrel.gov/pv/assets/pdfs/best-research-cell-efficiencies.
pdf (accessed on 10 November 2023).
6. Al-Ashouri, A.; Köhnen, E.; Li, B.; Magomedov, A.; Hempel, H.; Caprioglio, P.; Márquez, J.A.; Morales Vilches, A.B.; Kasparavicius,
E.; Smith, J.A. Monolithic perovskite/silicon tandem solar cell with >29% efficiency by enhanced hole extraction. Science 2020,
370, 1300–1309. [CrossRef]
7. Chen, B.; Zhengshan, J.Y.; Manzoor, S.; Wang, S.; Weigand, W.; Yu, Z.; Yang, G.; Ni, Z.; Dai, X.; Holman, Z.C. Blade-coated
perovskites on textured silicon for 26%-efficient monolithic perovskite/silicon tandem solar cells. Joule 2020, 4, 850–864. [CrossRef]
8. Hyun, J.Y.; Yeom, K.M.; Lee, S.-W.; Bae, S.; Choi, D.; Song, H.; Kang, D.; Hwang, J.-K.; Lee, W.; Lee, S. Perovskite/Silicon Tandem
Solar Cells with a Voc of 1784 mV Based on an Industrially Feasible 25 cm2 TOPCon Silicon Cell. ACS Appl. Energy Mater. 2022, 5,
5449–5456. [CrossRef]
9. Pyun, D.; Lee, S.-W.; Kim, Y.; Jang, G.S.; Choi, D.; Jeong, S.H.; Song, H.; Lee, S.; Cho, S.; Kim, J. First Demonstration of Top
Contact-Free Perovskite/Silicon Two-Terminal Tandem Solar Cells for Overcoming the Current Density Hurdle. ACS Appl. Energy
Mater. 2023, 6, 2687–2697. [CrossRef]
10. Pyun, D.; Choi, D.; Bae, S.; Lee, S.-W.; Song, H.; Jeong, S.H.; Lee, S.; Hwang, J.-K.; Cho, S.; Lee, H. Titanium Silicide: A Promising
Candidate of Recombination Layer for Perovskite/Tunnel Oxide Passivated Contact Silicon Two-Terminal Tandem Solar Cells.
ACS Appl. Mater. Interfaces 2024, 16, 28379–28390. [CrossRef] [PubMed]
11. Jeong, S.-H.; Hwang, J.-S.; Hwang, J.-K.; Lee, S.-W.; Lee, W.; Lee, S.; Pyun, D.; Cho, S.; Choe, Y.; Lee, H.-S. Potassium chloride
passivation for sputtered SnO2 to eliminate hysteresis and enhance the efficiency of perovskite solar cells. J. Alloys Compd. 2023,
968, 171890. [CrossRef]
Energies 2024, 17, 5998 11 of 12

12. Bae, M.-S.; Chang, Y.H.; Li, M.Q.; Kim, S.-W.; Lee, W.; Bae, J.-J.; Kim, D.; Kim, H.-S.; Lim, J.; Lee, J. Enhanced chemical interaction
between ionic liquid and halide perovskite to improve performance of perovskite solar cells. Mater. Today Energy 2024, 43, 101593.
[CrossRef]
13. Hwang, J.-K.; Jeong, S.-H.; Kim, D.; Lee, H.-S.; Kang, Y. A Review on Dry Deposition Techniques: Pathways to Enhanced
Perovskite Solar Cells. Energies 2023, 16, 5977. [CrossRef]
14. Kim, B.-S.; Pérez-del-Rey, D.; Paliwal, A.; Dreessen, C.; Sessolo, M.; Bolink, H.J. Simple approach for an electron extraction layer
in an all-vacuum processed nip perovskite solar cell. Energy Adv. 2022, 1, 252–257. [CrossRef] [PubMed]
15. Hwang, J.-K.; Lee, S.-W.; Lee, W.; Bae, S.; Kang, D.; Jeong, S.-H.; Lee, S.; Pyun, D.; Hwang, J.-S.; Cho, S. Sputtered PbI2 with
Post-Processing for Perovskite Solar Cells. Solar RRL 2023, 7, 2300214. [CrossRef]
16. Lee, J.; Kang, H.; Kim, G.; Back, H.; Kim, J.; Hong, S.; Park, B.; Lee, E.; Lee, K. Achieving large-area planar perovskite solar cells
by introducing an interfacial compatibilizer. Adv. Mater. 2017, 29, 1606363. [CrossRef]
17. Kim, S.-W.; Moon, S.J.; Kim, S.-H.; Yoo, J.J.; Kim, D.; Kim, B.-S.; Jeon, N.J. Reducing Humidity Dependency of Ambient-Air-
Processed Wide-Bandgap Inverted Perovskite Solar Cells. ACS Energy Lett. 2023, 8, 4777–4781. [CrossRef]
18. Yan, L.; Huang, H.; Cui, P.; Du, S.; Lan, Z.; Yang, Y.; Qu, S.; Wang, X.; Zhang, Q.; Liu, B. Fabrication of perovskite solar cells in
ambient air by blocking perovskite hydration with guanabenz acetate salt. Nat. Energy 2023, 8, 1158–1167. [CrossRef]
19. Zeng, Q.; Ma, Q.; Li, L.; Zheng, B.; Pan, Y.; Zhao, X.; Xiao, H.; Yan, C.; Liu, F. Tailoring particle size of PbI2 towards efficient
perovskite solar cells under ambient air conditions. Chem. Commun. 2023, 59, 5269–5272. [CrossRef] [PubMed]
20. Zheng, X.; Kong, W.; Wen, J.; Hong, J.; Luo, H.; Xia, R.; Huang, Z.; Luo, X.; Liu, Z.; Li, H. Solvent engineering for scalable
fabrication of perovskite/silicon tandem solar cells in air. Nat. Commun. 2024, 15, 4907. [CrossRef]
21. Ghosh, S.; Boro, B.; Porwal, S.; Mishra, S.; Singh, T. Role of antisolvent temperature and quaternary ammonium cation-based
ionic liquid engineering in the performance of perovskite solar cells processed under air ambient conditions. Energy Adv. 2023, 2,
1155–1165. [CrossRef]
22. Boro, B.; Mishra, S.; Singh, P.; Lahiri, B.; Varshney, S.K.; Singh, T. Thermal Stability Analysis of Formamidinium–Cesium-Based
Lead Halide Perovskite Solar Cells Fabricated under Air Ambient Conditions. Energy Technol. 2024, 12, 2400034. [CrossRef]
23. Mesquita, I.; Andrade, L.; Mendes, A. Effect of relative humidity during the preparation of perovskite solar cells: Performance
and stability. Solar Energy 2020, 199, 474–483. [CrossRef]
24. Berhe, T.A.; Su, W.-N.; Chen, C.-H.; Pan, C.-J.; Cheng, J.-H.; Chen, H.-M.; Tsai, M.-C.; Chen, L.-Y.; Dubale, A.A.; Hwang, B.-J.
Organometal halide perovskite solar cells: Degradation and stability. Energy Environ. Sci. 2016, 9, 323–356. [CrossRef]
25. Lee, S.-W.; Kim, S.; Bae, S.; Cho, K.; Chung, T.; Mundt, L.E.; Lee, S.; Park, S.; Park, H.; Schubert, M.C. UV degradation and
recovery of perovskite solar cells. Sci. Rep. 2016, 6, 38150. [CrossRef] [PubMed]
26. Hsu, R.-Y.; Liang, Y.-J.; Hung, Y.-J.; Lin, Y.-C. Impact of humidity in triple cation perovskite solar cells: Surface analysis. Mater. Sci.
Semicond. Process. 2022, 152, 107100. [CrossRef]
27. Purev-Ochir, B.; Liu, X.; Fujita, Y.; Semba, D.; Raju, T.B.; Tumen-Ulzii, G.; Wachi, A.; Sato, H.; Matsushima, T.; Adachi, C.
Oxygen-Induced Reversible Degradation of Perovskite Solar Cells. Solar RRL 2023, 7, 2300127. [CrossRef]
28. Cheng, Y.; Li, H.-W.; Qing, J.; Yang, Q.-D.; Guan, Z.; Liu, C.; Cheung, S.H.; So, S.K.; Lee, C.-S.; Tsang, S.-W. The detrimental effect
of excess mobile ions in planar CH3 NH3 PbI3 perovskite solar cells. J. Mater. Chem. A 2016, 4, 12748–12755. [CrossRef]
29. Manspeaker, C.; Merrion, G.; Krouse, K.; Zakhidov, A. Ambient processing conditions and their effects on perovskite device
performance. Cryst. Res. Technol. 2022, 57, 2100116. [CrossRef]
30. Toloueinia, P.; Khassaf, H.; Shirazi Amin, A.; Tobin, Z.M.; Alpay, S.P.; Suib, S.L. Moisture-induced structural degradation in
methylammonium lead iodide perovskite thin films. ACS Appl. Energy Mater. 2020, 3, 8240–8248. [CrossRef]
31. Howard, J.M.; Wang, Q.; Srivastava, M.; Gong, T.; Lee, E.; Abate, A.; Leite, M.S. Quantitative predictions of moisture-driven
photoemission dynamics in metal halide perovskites via machine learning. J. Phys. Chem. Lett. 2022, 13, 2254–2263. [CrossRef]
[PubMed]
32. Srivastava, M.; Hering, A.R.; An, Y.; Correa-Baena, J.-P.; Leite, M.S. Machine learning enables prediction of halide perovskites’
optical behavior with >90% accuracy. ACS Energy Lett. 2023, 8, 1716–1722. [CrossRef] [PubMed]
33. Mammeri, M.; Bencherif, H.; Dehimi, L.; Hajri, A.; Sasikumar, P.; Syed, A.; AL-Shwaiman, H.A. Stability forecasting of perovskite
solar cells utilizing various machine learning and deep learning techniques. J. Opt. 2024, 1, 1–9. [CrossRef]
34. Ko, Y.; Lee, C.; Kim, Y.; Kim, Y.; Yun, Y.J.; Jun, Y. Dew point temperature as an invariant replacement for relative humidity for
advanced perovskite solar cell fabrication systems. J. Mater. Chem. A 2018, 6, 20695–20701. [CrossRef]
35. Kim, S.; Bae, S.; Lee, S.-W.; Cho, K.; Lee, K.D.; Kim, H.; Park, S.; Kwon, G.; Ahn, S.-W.; Lee, H.-M. Relationship between ion
migration and interfacial degradation of CH3 NH3 PbI3 perovskite solar cells under thermal conditions. Sci. Rep. 2017, 7, 1200.
[CrossRef] [PubMed]
36. Bae, S.; Kim, S.; Lee, S.-W.; Cho, K.J.; Park, S.; Lee, S.; Kang, Y.; Lee, H.-S.; Kim, D. Electric-field-induced degradation of
methylammonium lead iodide perovskite solar cells. J. Phys. Chem. Lett. 2016, 7, 3091–3096. [CrossRef] [PubMed]
37. Burst, J.M.; Duenow, J.N.; Albin, D.S.; Colegrove, E.; Reese, M.O.; Aguiar, J.A.; Jiang, C.-S.; Patel, M.; Al-Jassim, M.M.; Kuciauskas,
D. CdTe solar cells with open-circuit voltage breaking the 1 V barrier. Nat. Energy 2016, 1, 16015. [CrossRef]
38. Scarpulla, M.A.; McCandless, B.; Phillips, A.B.; Yan, Y.; Heben, M.J.; Wolden, C.; Xiong, G.; Metzger, W.K.; Mao, D.; Krasikov, D.
CdTe-based thin film photovoltaics: Recent advances, current challenges and future prospects. Sol. Energy Mater. Sol. Cells 2023,
255, 112289. [CrossRef]
Energies 2024, 17, 5998 12 of 12

39. Kartopu, G.; Oklobia, O.; Turkay, D.; Diercks, D.; Gorman, B.; Barrioz, V.; Campbell, S.; Major, J.; Al Turkestani, M.; Yerci, S. Study
of thin film poly-crystalline CdTe solar cells presenting high acceptor concentrations achieved by in-situ arsenic doping. Sol.
Energy Mater. Sol. Cells 2019, 194, 259–267. [CrossRef]
40. Ablekim, T.; Colegrove, E.; Metzger, W.K. Interface engineering for 25% CdTe solar cells. ACS Appl. Energy Mater. 2018, 1,
5135–5139. [CrossRef]
41. Wang, Y.; Dar, M.I.; Ono, L.K.; Zhang, T.; Kan, M.; Li, Y.; Zhang, L.; Wang, X.; Yang, Y.; Gao, X. Thermodynamically stabilized
β-CsPbI3 –based perovskite solar cells with efficiencies >18%. Science 2019, 365, 591–595. [CrossRef]
42. Chen, H.; Fu, W.; Huang, C.; Zhang, Z.; Li, S.; Ding, F.; Shi, M.; Li, C.Z.; Jen, A.K.Y.; Chen, H. Molecular engineered hole-extraction
materials to enable dopant-free, efficient p-i-n perovskite solar cells. Adv. Energy Mater. 2017, 7, 1700012. [CrossRef]
43. Chouhan, A.S.; Jasti, N.P.; Avasthi, S. Effect of interface defect density on performance of perovskite solar cell: Correlation of
simulation and experiment. Mater. Lett. 2018, 221, 150–153. [CrossRef]
44. Cai, Q.; Lin, Z.; Zhang, W.; Shen, G.; Wen, X.; Dong, H.; Xu, X.; Zhu, D.; Mu, C. Improvement performance of planar perovskite
solar cells by bulk and surface defect passivation. ACS Sustain. Chem. Eng. 2021, 9, 13001–13009. [CrossRef]
45. Burgelman, M.; Nollet, P.; Degrave, S. Modelling polycrystalline semiconductor solar cells. Thin Solid Film. 2000, 361, 527–532.
[CrossRef]
46. Abnavi, H.; Maram, D.K.; Abnavi, A. Performance analysis of several electron/hole transport layers in thin film MAPbI3 -based
perovskite solar cells: A simulation study. Opt. Mater. 2021, 118, 111258. [CrossRef]
47. Grinsztajn, L.; Oyallon, E.; Varoquaux, G. Why do tree-based models still outperform deep learning on typical tabular data? Adv.
Neural Inf. Process. Syst. 2022, 35, 507–520.
48. Scott, M.; Su-In, L. A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst. 2017, 30, 4765–4774.
49. Lundberg, S.M.; Erion, G.; Chen, H.; DeGrave, A.; Prutkin, J.M.; Nair, B.; Katz, R.; Himmelfarb, J.; Bansal, N.; Lee, S.-I. From local
explanations to global understanding with explainable AI for trees. Nat. Mach. Intell. 2020, 2, 56–67. [CrossRef]
50. Lee, M.-H. Interpretable machine-learning for predicting power conversion efficiency of non-halogenated green solvent-processed
organic solar cells based on Hansen solubility parameters and molecular weights of polymers. Sol. Energy 2023, 261, 7–13.
[CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.