Energy and AI 10 (2022) 100203

Contents lists available at ScienceDirect

Energy and AI
journal homepage: www.sciencedirect.com/journal/energy-and-ai

Visualization-based prediction of dendritic copper growth in

electrochemical cells using convolutional long short-term memory
Roshan Kumar a, Trina Dhara a, Han Hu b, *, Monojit Chakraborty a, *
a
Department of Chemical Engineering, Indian Institute of Technology Kharagpur, Kharagpur, 721302, India
b
Department of Mechanical Engineering, University of Arkansas, Fayetteville, AR, 72701, United States of America

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• Metallic dendrites in the charging phase

of batteries cause performance
degradation.
• Real-time imaging of copper electrode­
posits to mimic the randomized metallic
growth.
• Convolutional Long Short-Term archi­
tecture-based analysis growth
prediction.
• Pixel level assessments between the
experimental and predicted images.
• Foreseeing the morphology without
even performing the experiments.

A R T I C L E I N F O A B S T R A C T

Keywords: Electrodeposition in electrochemical cells is one of the leading causes of its performance deterioration. The
Electrodeposition prediction of electrodeposition growth demands a good understanding of the complex physics involved, which
Electrochemical cell can lead to the fabrication of a probabilistic mathematical model. As an alternative, a convolutional Long short-
Deep learning
term memory architecture-based image analysis approach is presented herein. This technique can predict the
Data-driven modelling
Convolutional long short-term memory
electrodeposition growth of the electrolytes, without prior detailed knowledge of the system. The captured
images of the electrodeposition from the experiments are used to train and test the model. A comparison between
the expected output image and predicted image on a pixel level, percentage mean squared error, absolute per­
centage error, and pattern density of the electrodeposit are investigated to assess the model accuracy. The
randomness of the electrodeposition growth is outlined by investigating the fractal dimension and the interfacial
length of the electrodeposits. The trained model predictions show a significant promise between all the exper­
imentally obtained relevant parameters with the predicted one. It is expected that this deep learning-based
approach for predicting random electrodeposition growth will be of immense help for designing and opti­
mizing the relevant experimental scheme in near future without performing multiple experiments.

1. Introduction coating and modification of the surface properties of metal [1]. It is one
of the key processes used in various engineering [2] and scientific ap­
Electrolytic deposition is one of the conventional processes for the plications [3]. The method involves the formation of solid metal deposits

* Corresponding authors.
E-mail addresses: [email protected] (H. Hu), [email protected] (M. Chakraborty).

https://doi.org/10.1016/j.egyai.2022.100203

Available online 17 September 2022

2666-5468/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
R. Kumar et al. Energy and AI 10 (2022) 100203

from an ionic solution on the application of an external electric field [4]. complex output results from a set of unknown input data. Recently Attia
The electrodeposition technique is extensively studied for past decades et al [32]. have developed a machine learning methodology to predict a
to explore different morphological variations of the metallic deposits parameter space for maximizing battery cycle life, where they have
[5].The deposits produced by this process can produce a large variety of shown, how this methodology can reduce the optimization cost and time
fascinating shapes [6]. Among all, the tree-like, self-similar, repetitive for experiments by foretelling the final cycle life of the battery. Apart
branched structure observed during the electrodeposition [7], is known from that, Ng et al [33]. have provided a wide overview of machine
as dendritic growth. This is also one of the major concerns pertaining to learning algorithms to predict the state of charge (SOC) and state of
the generation of the consistent energy supply of an electrochemical cell. health (SOH) for the battery management system. Tu et al [34]. used an
This uncontrolled growth is highly undesirable in rechargeable batteries artificial neural network with a physics-based model to predict the state
involving metal anodes [8]. Lithium-ion batteries are extensively stud­ of lithium-ion batteries. Machine learning techniques are proved to be
ied as a means of energy storage in diverse electronic devices in the past very promising even for predicting the deposition thickness of ZnNi
decades [9]. Conventional Lithium-ion batteries have Lithium interca­ electroplating [35]. Electrodeposition is an integral part of a battery
lated transition metal oxide cathodes and graphite anodes. The charging (during the charging cycle). The morphological variation of the elec­
phase involves the application of an external electric field, and the trodeposits during the charging cycle is completely random and un­
Li-ions get transported to the graphite anode. Therefore, electrodepo­ predictable. Extensive research is still going on to capture the variety of
sition is the fundamental process to carry out the charging phase in a such outgrowing structures during the electrodeposition. This will help
rechargeable battery. Similarly, during discharge, the Li ions again in understanding the dynamic behavior of the deposited growth as well
travel from the graphite electrode to the cathode material. The con­ as the heterogeneity of the structural evolution [36]. The use of the
ventional graphite-based anode is replaced by Lithium metal due to its real-time experimental frames of growing electrodeposits to predict the
high specific capacity (3860 mAhg− 1). This resulted in a major opera­ dynamics of the deposition is inadequate. Forecasting the deposition
tional drawback due to the formation of metallic dendrites during the dynamics is tremendously advantageous for experimentalists working in
charging of metal anode-based batteries [10, 11]. These self-repeating the battery industries. For example, with a given set of initial data, the
growing metallic units may form a bridge connecting the two elec­ time at which the growing unit touches the other electrode to cause
trodes after several charging cycles, leading to a short circuit. Another short-circuit may be foreseen even without performing a trial. This can
factor is the formation of Solid Electrolyte Interface (SEI) [12], which is prevent unwanted damage to the entire circuit. However, a model
an electrically insulating layer that may form when the electrolyte capable of predicting metal electrodeposition is scant in literature. A
molecules come in contact with the lithium metal. SEI is capable of physical model depends upon a lot of measured parameters to predict
passing Li ions but incapable of conducting electricity. Thus, this insu­ electrolytic growth [11]. Several phase-field based theoretical simula­
lating layer covers freshly deposited solid Lithium on the metal anode. tions [37] and mathematical models [38] are present in the literature.
Sometimes, a dendrite can break down at its weaker neck. As a result of Most of the models deal with the physics involved and also include
that, as soon as the Li atoms at the cross-section of the neck are exposed various variables which make the prediction uncertain. Generally, these
to the electrolyte molecules, they get covered with the SEI layer. Thus an models deal with the physics involved behind this random uncontrolled
insulated dendritic structure (‘dead Lithium’) is formed [13], which is growth. Unlike physical models, a machine learning model can work
no longer a part of the electrode surface [14]. This insulated broken part with simple input like images, which is highly advantageous in pre­
is known as Columbic loss [15]. The ‘dead Lithium’ accumulated over dicting the dynamics of a rapidly changing system like self-driving cars
the electrode surface and may cause extra resistance for any incoming [39]. For example, XGB algorithm [35] is observed to predict electro­
Li-ion to participate in the electrodeposition process. This may lead to deposition thickness with a high correlation coefficient. Herein, the
heat generation and even thermal run away[16]. In this context, it is application of the machine learning concept has been shown using
worth paying attention to the other metal anodes like Cu [17], Na [18], real-time experimental frames of growing electrodeposits. It is to be
Zn [19], Ag [20], and Mg [21], which also can procedure dendritic noted that most of the metal anodes (Li, Na, Mg, etc.) used in the battery
deposition in the charging cycle. Investigation of copper electrodeposits industry are highly reactive, and cannot be handled at room temperature
are observed to become the pivot for decades. Schilardi et al [17]. without an inert atmosphere [18,21,40]. As the objective is to predict
investigated the in situ evolution of the growing interface of copper, and the deposition dynamics by image-based ML techniques, a safer metal
the results are validated with the Edward-Wilkinson equation [22]. electrode has been chosen so that a pool of real-time images can be
Schneideret al [23]. investigated the growth front behavior of copper captured easily with clarity. Therefore, to mimic the real scenario, a
electrodeposition using liquid cell TEM by image analysis technique. model Copper-Copper symmetric electrochemical cell has been
Several other experimental studies have also been performed to observe employed due to its availability and ease of handling. Investigation of
the morphological variation of copper electrodeposits at different copper electrodeposits are observed to become the pivot for decades.
overpotentials [24]. It is observed that mostly video microscopy [25] Schilardi et al [17]. investigated the in situ evolution of the growing
and post-experiment image analysis [26,27] are used to determine the interface of copper, and the results are validated with the
nature of the deposition growth pattern [28]. As a result of that, Edward-Wilkinson equation [22]. Schneideret al [23]. investigated the
monitoring of the deposition dynamics is strictly limited to the experi­ growth front behavior of copper electrodeposition by analysing images
mental time scale and other complex instrument availability. However, obtained from liquid cell transmission electron microscopy (TEM) .
detailed investigations of the effect and the dynamics of the electrode­ Several other experimental studies have also been performed to observe
position growth beyond the experimental duration are challenging and the morphological variation of copper electrodeposits at different
limited in the literature. Machine Learning (ML) techniques can be overpotentials [24].
utilized to confront this problem and that may revolutionize electro­ In this study, the real-time frames of electrodeposition on the copper
chemical science [29]. electrode are recorded using a video microscope at 250 frames/s. With
Machine learning is a branch of artificial intelligence (AI), that al­ the given set of initial data obtained from two different experiments, a
lows a data-driven prediction even without the requirement of the single model is proposed, that can successfully predict the deposition
physical insights of the process [30,31]. Electrodeposition is a complex dynamics for different concentration ranges. It must be stated here that
process that depends on many system variables. To model such complex the choice of the metal electrode has significant effect on the electro­
processes, machine learning tools along with process physics are proven deposition mechanism [41]. For instance, a specific metal electrode has
to be helpful. Deep learning and AI-based machine learning methods are unique electrode-electrolyte interaction, self-diffusion barrier [42], and
capable of fitting high-dimensional data or system variables in a mechanical properties of the SEI, which affect the deposition charac­
generalized model. The model can be used later for the prediction of teristics [43]. The proposed model herein solely operates with the

2
R. Kumar et al. Energy and AI 10 (2022) 100203

captured images as input to it. It is a data-driven model, which can be 2. Methodology

further improved to understand the complex phenomena of electrode­
position leveraging AI techniques. The model does not take into account 2.1. Experiments for Data-set preparation
the material property or the different types of intricate structural vari­
ations of different metals. Therefore, it is worth claiming that, although An electrochemical cell is prepared using a transparent plastic
the deposition dynamics of the present study is restricted to copper cuvette. A schematic of the cell is depicted in Fig. 1 and a snapshot of the
electrodeposition only, this model can be used for a wider perspective, to entire setup is presented in Supporting Information S1 (Figure S1). The
predict the growth behavior on the other metal electrodes as well. experimental specifications are followed using Dhara et al [44]. The
Conclusively, electrodeposition prediction using a general mathematical setup consists of two electrodes made of bare copper wires (diameter, d
model is strenuous and computationally expensive. The proposed deep = 2.019 × 10− 4 m). The copper electrodes are cleaned by rinsing in
learning model is an image analysis-based model, and the same model acetone, isopropyl alcohol, and deionized water to remove surface
can be trained for different experimental conditions. A data-driven contaminants. The cleaned copper wires are dipped into the electrolyte
model like this model may help in anticipating the battery’s health and a constant inter-electrode gap (δ = 5 × 10− 3 m) is maintained.
without performing multiple experiments during the design of the It is to be noted that the electrochemical deposition depends on
electromedical cells. All the features of this image-based deep learning several factors such as the applied potential difference, inter-electrode
model prepared herein, using a different approach, make it unique gap, material and shape of the electrode, electrolyte property, addi­
compared to the other existing models in the literature. The success of tives, operating temperature, etc. For the present study, only two influ­
this AI-based model prediction will certainly help in improving the encing factors are considered, namely the concentration of electrolyte
present energy storage related issues. solution and the deposition time. Other parameters are made constant
for all the experiments. The electrolyte in the present study is prepared
by diluting analytical grade copper sulfate pentahydrate (CuSO4,5H2O)
in deionized water. Electrolyte solutions of 0.025 M and 0.1 M CuSO4
are prepared for conducting two different sets of experiments. The
prepared electrolyte solutions are then filled into the cuvette. From the

Fig. 1. Schematic representation of the electrochemical cell modified and reprinted from Dhara et al [44]. Two copper wires serve as electrodes. On the application
of voltage, the deposition of metallic copper takes place on the negative electrode. The schematic portrayal of the dendritic electrodeposits is also presented.
Figure not to scale.

3
R. Kumar et al. Energy and AI 10 (2022) 100203

previous study [44], a comparatively uniform deposition was observed 0 to255. These images are normalized and the final pixel value is set in
at a higher electrolyte concentration as compared to the branched the range of 0 to1. Further binary segmentation is performed so that the
dendritic structure at a lower one. Hence, the two representative elec­ pixel values are either 0 or 1. The normalization and segmentation steps
trolyte concentrations (0.025 M and 0.1 M) are prepared herein to ensure fast training time. For a given sequence of images, vectors are
observe different morphological variations of the electrodeposition generated from the pixel value of images. After the data preprocessing
growth. After inserting the wires in the electrolyte solution, the cuvette step, an appropriate model for sequence learning has to be selected. An
is sealed with parafilm. A potential difference of 6 V is applied to the LSTM-based model is generated with assigned weight to the model. The
electrodes. The entire setup is placed in line with the camera to capture model is then trained and the binary cross entropy loss is monitored.
real-time images of metallic growth. During the training, the model weight is changed to optimize the loss
With the increase in the electrolyte concentration, the nature of the function. In the subsequent iterations, when the loss is reduced, the
dendritic growth changes from highly fractal to dense agglomerated weight is saved. If the weight remains constant through a few predefined
clusters [45]. The growing micro-structures consist of several morpho­ numbers of iterations, or if a maximum number of iterations is reached,
logical variations, such as dendritic, cauliflower-like, mossy, etc. How­ the training is stopped.
ever, all of them are lumped considering the randomness, and hereafter The current model consists of 4 stacked convolutional LSTM [51]
termed as dendritic growth. The observation window in the experiments (convLSTM) followed by a batch normalization [52] layer. The model
is kept fixed such that the growing deposits do not cross the vertical parameters like filters and kernel size are adjusted by monitoring the
frame dimension (240 pixels) with time. A set of 5566 images for 0.1 M loss and optimizing the computational cost. Processed image vectors of
and 5905 images for 0.025 M electrolyte concentration is obtained from dimension 100 × 80 × 128 × 1 are fed as input to the model. The input is
the experiments. A sample sequence of the images is presented in Fig. 2a. followed by a 2D–convolutional LSTM layer. Contrary to a normal
All the captured images have a dimension of 640 × 240 px. To reduce LSTM, a convLSTM uses convolutional transform in recurrent and input
the computational cost, all the images are converted (Fig. 2b) to binary transition, which facilitates feature extraction from the images and
images using the Chan-Vese segmentation algorithm [46]. training on 2D image data [53]. A schematic representation of the model
To inspect the presence of redundant and repetitive images in the architecture is depicted in Fig. 4.
sequence, the overall black pixel is calculated for each image. Images of To keep the vector dimensions uniform after the convolution oper­
similar nature, where the consecutive frame is indistinguishable with no ation, the padding is kept the same throughout the model. Padding in­
significant change in the electrodeposition pattern are removed. This serts a zero in the output vector so that the dimension is consistent with
ensures the model learns the electrodeposition growth pattern and not the input vector. The convLSTM uses tanh as the default activation
the experimental anomaly. The images obtained from the experiments function, which can take large inputs and outputs values in the range of
from both concentrations of electrolytes are used to train the model. [-1,1]. All LSTM units have 40 filters, with a 3 × 3 kernel size. The kernel
All the images are sorted and the image dimension is reduced to 80 × is responsible for extracting the feature map from the image vector. The
128 pixels to decrease the computational cost and speed up the training stride and dropout values are kept as default to avoid the down-sampling
process. The images are stacked and the input and output image vectors of the feature map. Each convLSTM is followed by a batch normalization
of the same dimension are generated. The output vector represents the layer, which helps to deal with the problem of “internal covariate shift”
next sequence of images for a given input image vector. The input and during the training. The batch normalization layer standardizes the
output vectors are shuffled and divided randomly into a train-test set in layer input and helps the model converge faster. The model has a 3D
the ratio of 80:20. convolution layer attached at the end, which extracts important infor­
mation from the output of convLSTM layers. The 2D convolution layer
performs a spatial convolution operation using a 2D kernal resulting in a
2.2. Model architecture 2D output, in contrast to a 3D convolution layer, which uses a 3D
cuboidal kernal and produces a 3D output. Using a 3D convolution layer
In the current study, the aim is to understand the future growth on a sequence of 2D images, the temporal feature variation in the input is
image sequence given the current growth pattern. The electrodeposition also captured. The 3D convolution layer is more suitable to extract
growth problem can be fit into the many to many sequence prediction spatiotemporal features from data [54] as it can perform convolution
categories. Recurrent neural networks (RNN) are well known for operations in 3D. Sigmoid activation is used in the 3D convolution layer,
sequence prediction and used for machine translation [47], speech and an output image vector having a pixel value between [0,1] is
recognition [48], time-series forecasting [49], protein structure pre­ obtained.
diction [50], etc. Basic RNN has an inherent problem of vanishing and The model uses binary cross-entropy loss as a performance matrix
exploding gradient, making it incompatible for training on a large (Supporting Information S2 (Figure S2) as it is easy to compare the
dataset. A special type of RNN such as gated recurrent unit (GRU) and predicted pixel value with an actual pixel value of the binary image. The
long short-term memory (LSTM) provide the solution to extreme model is optimized using the ‘adadelta’ optimizer [55]. Three different
gradient problems by leveraging the information flow control using a callback methods, namely ‘ModelCheckpoint’, ‘TimeHistory’, and ‘Ear­
gating mechanism. LSTM has a ‘forget’, ‘input’, and ‘output’ gate lyStopping’ are defined for the model. The ‘ModelCheckpoint’ is used to
structure which decides the important information to be carried in the save the best model weight. The ‘EarlyStopping’ is used with patience
network. value 7. Early Stopping can stop training the model if the performance
Fig. 3 represents the overall process, from collecting the experi­ matrix shows no improvement for a certain number of epochs. The
mental images to model training. The study starts by capturing experi­ TimeHistory callback records the history object and determines the time
mental images of electrodeposition growth on copper electrodes. The for each epoch. The model is fitted with a validation split of 0.05 for 50
images are grayscale and have a pixel intensity value in the range of

Fig. 2. (a) Real-time images of the growing electrodeposit for 0.025 M electrolyte concentration (b) Binary images obtained after Chan-Vese segmentation.

4
R. Kumar et al. Energy and AI 10 (2022) 100203

Fig. 3. Flowsheet of convolutional LSTM-based model training prepared using the experimentally obtained image dataset and used for the prediction of the elec­
trodeposition growth.

epochs on the bridges 2 GPU partition. A schematic of the entire algo­ difference of the pixel intensity value as presented in Eq. (1) where, Aij
rithm is presented in Fig. 5 for an improved understanding. and Bij represent pixel values at coordinate i, j in the compared images of
A and B. Black pixel and white pixel values have gray values of 0 and
3. Results and discussion 255.
( ) ( ) (∑ i=P ∑j=Q
)
100 1
Once the model is prepared, it is then trained on the electrodeposi­ %MSE = × × 2 2
Aij − Bij (1)
65025 P×Q
tion growth image dataset obtained for 0.025 M and 0.1 M electrolytic i=0 j=0

concentrations. The electrodeposition growths for the different con­

The square difference between the white and black pixels of the
centration ranges are observed to have different patterns. The model is
images used herein is 65025. When MSE of a completely black and white
expected to predict the growth pattern for different concentration ranges
image is performed, the summation of the square of pixel difference, and
with reasonable accuracy. A comparison of the real-time dynamic
an average of all the pixel values, is translated to 65025. The mean
electrodeposition growth for 0.025 M and 0.1 M electrolytic concen­
square error between all the experimental and predicted image is
tration is presented in the supporting information (Supporting Video S3,
determined using the mse function from the skimage library. Skimage is
S4). A sequence of the selected captured and model-predicted images is
a popular image processing and computer vision library [58]. The mse
presented in Fig. 6. After the deep learning model is trained and predicts
function computes the sum of the square of each pixel value and per­
results for input, it is important to validate the model prediction using
forms an average. The percentage MSE is calculated using the formula
appropriate comparison metrics. The comparison matrix can be a user-
presented in Eq. (1). MSE provides a pixel-level comparison between
defined or system-defined function that can capture all the intricacy of
two images.
the model prediction. In the current study, several metrics like per­
Figs. 7(a) and 7(b) presents the percentage mean square difference
centage mean square error, absolute percentage error, pattern density,
for 0.025 M and 0.1 M electrolyte-concentration respectively. The per­
interfacial length, and fractal dimension are used to establish the ac­
centage MSE is below 2.5% for the whole range of images which dem­
curacy and reliability of the model.
onstrates a fair resemblance between experimental and predicted
images. In the case of 0.1 M electrolyte concentration, the MSE is found
3.1. Percentage mean square error to be below 1% for most of the compared images. The average of the
percentage mean square error for 0.025 M and 0.1 M electrolytic con­
Mean square error (MSE) is one of the widely accepted image com­ centrations are found to be 2 and 0.2% respectively. The uniform growth
parison metrics in the computer vision field [56,57]. For an image of (see Supporting Video S4) can be a probable cause for lower MSE in 0.1
dimension P × Q, the mean square error is defined as the mean square M electrolyte concentration as compared to the other experiment with

5
R. Kumar et al. Energy and AI 10 (2022) 100203

Fig. 4. Model architecture of the convolutional LSTM (a) with corresponding input-output vector size across the different layers of the model (b) with layers
including the number of filters, kernel size, padding, and activation function details.

0.025 M electrolyte concentration. ratio [59] of the growing electrodeposits, which can be one of the other
physical properties of the electrodeposits is not considered herein. The
reason behind this is essentially the real-time 2D images captured from
3.2. Absolute percentage error the experiments. In the experimentally obtained 2D real-time images,
the number of pores captured (white pixels surrounded by black pixels)
After comparing the pixel level correlation between the experiment inside the intricate compact structures of growing deposits is likely to be
and predicted images, the percentage difference in the experimental and erroneous. As the images are 2D and a growing unit that is exactly rising
predicted image area is calculated. The area of any image is determined behind another dendritic unit, will block the visibility of the pores of the
by counting the black pixel of the image. The absolute percentage error structure at the front. To characterize this exceedingly random struc­
metric represents the percentage dissimilarity between the area of the tural progression, the definition provided by Saab and Sultan [60] is
predicted and experimental image. Mathematically the absolute per­ followed to determine another parameter, hereafter termed Pattern
centage error is represented in Eq. (2) as, Density.
⃒ ⃒
⃒Areaexperimental − Areapredicted ⃒
Absolute percentage error = 100 × ⃒⃒ ⃒
⃒ (2)
Areaexperimental
3.3. Pattern density
Fig. 8 represents the variation of the absolute error (%) for different
concentrations of electrolyte with time. At the beginning of the experi­ Pattern Density is defined as the percentage area covered by the
ment, when the dendritic area is less due to limited deposition, a rela­ electrodeposits (number of the black pixel as the percentage of the total
tively higher absolute error percentage is observed for both the number of pixels of the image concerned). Therefore, the transient
electrolyte concentration. With the increase in the dendritic area as the variation of pattern density (%) depicts the kinetic evolution of the
experiment proceeds, the absolute error percentage declines. This growing front. Once the model is trained, it is used to predict the output
demonstrates that the model prediction becomes extremely precise as sequence made from the input sequence of all the experimental images.
the experiment progresses. The relatively less accurate prediction at the The predicted images can be compared with the next sequence of
very early stage of the experiment is may be attributed due to the smaller experimental images to get the accuracy of the model. Predicting the
electrodeposition area and higher growth rate. Overall, it is observed whole sequence helps to visualize the electrodeposition growth area
that the growth rate is higher during the initial phase [44], and the with time. The pixel area measures the similarity between two images
dynamic evolution of metal deposits is highly irregular. The mesopore based on an aggregated value. The detailed calculation of pattern

6
R. Kumar et al. Energy and AI 10 (2022) 100203

Fig. 5. Algorithmic representation of the process from experimental images to predicted images for the prediction of near future electrodeposition growth using the
convolutional LSTM model.

Fig. 6. Image sequence of the electrodeposition (a) From experiment for 0.025 M electrolyte concentration; (b) model predicted image sequence with time for 0.025
M electrolyte concentration.

density is presented in Supporting Information S5. between the pattern density of the experimental and predicted images
Fig. 9(a) and 9(b) represent the percentage of experimental and for both 0.025 M and 0.1 M electrolyte concentrations is found to be less
predicted electrodeposition growth for different concentrations. than 1, as presented in Supporting Information S6 (Figure S3). The
Initially, the electrodeposition growth shows a rapid increase, and with average error in pattern density percentage prediction for 0.025 M and
time a gradually decreasing trend is observed. With the increase in the 0.1 M electrolytic concentration is found to be 2.3 and 1.3% respec­
electrolytic concentration, the grain size of the deposit rises, and a tively. The pattern densities for predicted and experimental images for
transition from fractal to bulk shape crystal is observed [21]. Due to the both concentrations show a similar trend and are barely distinguishable.
large crystal size in 0.1 M electrolyte concentration, the growth appears This shows the capability of the model to perfectly learn the underlying
to be uniform and changes gradually as compared to the occurrence with recurrent relationship of temporal growth of dendrites with concentra­
0.025 M electrolyte concentration. The root mean square error (RMSE) tion as a parameter. The pattern density shows a similar trait as found by

7
R. Kumar et al. Energy and AI 10 (2022) 100203

Fig. 7. Percentage mean square error by comparing the images obtained from the experiments with the model predicted one, for (a) 0.025 M, and (b) 0.1 M
electrolyte concentration.

Fig. 8. Percentage absolute error by comparing the calculated area obtained from the experimentally obtained images with the model predicted one for (a) 0.025 M
electrolyte concentration; (b) 0.1 M electrolyte concentration. (BP: Number of black pixels, WP: Number of white pixels).

Fig. 9. Pattern density percentage obtained by comparing the black pixel percentage, for the electrolyte concentration of (a) 0.025 M and (b) 0.1 M.

8
R. Kumar et al. Energy and AI 10 (2022) 100203

dendritic peak height calculation, using molecular simulation [61]. It dimension is considered herein to further comprehend the fluctuations
must be stated here that though pattern density successfully determines of the electrodeposits. Fractal dimension is a commonly used method to
the 2D deposition area of a specific frame, the characterization of the represent torturous surfaces [64].
surface undulation over the time experimental span (Fig. 2, from flat
interface to extremely irregular topology) is unresolved. Therefore, to
explore it, the transient variation of the length of the growing interface is 3.5. Fractal dimension
calculated and hereafter indexed as the interfacial length.
The fractal dimension is based on the idea that the detail of a pattern
varies when the scale of measurement changes. The fractal dimension
3.4. Interfacial length can be understood as the space-filling capacity of any pattern. Hence, it
is important to understand the electrodeposition growth [65].
The electrodeposition results in uneven growth in random directions. Geometrically, fractals are irregular complex structures that may have
The interfacial length can be defined as the overall length of the repeating substructures. Fractal dimension (D) is used to characterize
enclosure surrounding the electrodeposit growth interface (see Sup­ the complexity of a fractal, where complexity depends upon surface
porting Information S7 (Figure S4) for details). To calculate the inter­ detail and scale of measurement. In a fractal having N self-similar
facial length, images obtained from the experiments are binary shapes, and a scaling factor of ε, the fractal dimension D is defined
segmented using a threshold value of 127 using the OpenCV (cv2) and presented in Eq. (3) as,
threshold function. This resulted in a conversion of the grayscale image
logN
to a binary image. All the contour hierarchies in the image are found logNε = − D = (3)
log ε
using the findContour function using chain approx as a parameter [62].
FindContours function detects the change in image color and marks the Various approaches like the Richardson method [66] and the
point as a contour point. Chain approx is a way to store the coordinates box-counting method [67] are used to calculate the fractal dimension.
of the endpoints among all contour points which are in a straight line. The box-counting method is used to find the fractal dimension of com­
The perimeter of the contour is found using the arcLength [63] function plex shapes and is currently employed in this study. In the box-counting
in OpenCV. The arcLength function takes the points which are given out method, 2 grids of different scales are used and the number of grid boxes
by the findContour function and calculate the distance between containing the image is counted [68]. For the grid counting method
consecutive contour points as the sum of the square of the coordinates of using 2 scales εa and εb, the fractal dimension is presented in Eq. (4) and
the points. defined as [68]
An interfacial length variation of the electrodeposits with time, is
logNa − logNb
presented in Fig. 10. For 0.1 M electrolyte concentration (Fig. 10(b)), the − D= (4)
logεa − logεb
interface length increases continuously and reached a maximum value
with time. It is observed that for 0.025 M electrolyte concentration Where Na and Nb are the numbers of the box containing the image.
(Fig. 10(a)), the interfacial length increases with time, and then reaches Fig. 11(a) depicts the fractal dimension for the experimental and
a peak near 5.5 s. This peak represents maximum random electrodepo­ predicted images obtained for the electrodeposition growth at 0.025 M
sition growth. From the figures in Fig. 10, it is clear that the deposition electrolyte concentration. The fractal dimension calculated from the
results in tree-like structures which are responsible for high interfacial images increases and reaches a maximum before starts declining, which
length. With time, the tree-like structure merges, resulting in a decrease suggests that the electrodeposition rises and has a highly irregular
in the interfacial length as can be observed from the experimental and structure during the electrochemical reaction. The fractal dimension
predicted images in the inset. The average RMSE between the experi­ shows an irregular trend after 10 s, which may be attributed to the
mental and predicted images are found to be 44.4 and 30.3 for 0.025 M inconsistent electrodeposition growth. Fig. 11(b) presents the fractal
and 0.1 M electrolyte concentrations respectively. A more detailed dy­ dimension for the electrodeposition growth at 0.1 M electrolyte con­
namic analysis of the RMSE is presented in Supporting Information S7 centration. Initially, it shows an overall rising trend followed by a minor
(Figure S5). The average error in the prediction of interfacial lengths is change as the experiment progresses. This suggests that the electrode­
observed to be 10.7 and 9.2% for 0.025 M and 0.1 M electrolytic con­ position growth in 0.1 M concentration is more consistent as compared
centration respectively. Though the interfacial length perfectly de­ to 0.025 M concentration. The model predicts the fractal dimension with
scribes the interface fluctuations, another parameter namely the fractal a 2.9% and 4.3% average deviation from the experimentally obtained

Fig. 10. Relative comparison of the length obtained from the experiments and the model prediction by calculating the length of the total interface of the elec­
trodeposits for (a) 0.025 M electrolyte concentration and (b) 0.1 M electrolyte concentration.

9
R. Kumar et al. Energy and AI 10 (2022) 100203

Fig. 11. Relative comparison of the fractal dimension obtained from the experiments and the model predicted image sequences, for the electrolyte concentration of
(a) 0.025 M and (b) 0.1 M.

fractal dimension for 0.025 M and 0.1 M electrolyte concentration, architecture consisting of convolutional long short-term memory is
respectively. explored, which is capable of predicting the spatio-temporal variation of
It must be further noted that multiple parameters related to elec­ dendrites. The model is trained on two different datasets with varying
trodeposition are investigated by considering two different electrolyte concentrations. Various metrics like percentage MSE, pattern density,
concentrations (0.025 M and 0.1 M) only. The primary cause [69] of the interfacial length, and absolute percentage error are determined. During
growth of the electrodeposits is due to the combined effect of the testing, the mean squared error and the percentage absolute error are
electro-migration and diffusion on the reactant ions (in absence of found to be below 3% for all model predicted outcomes. A fast-growing
convection). The transport of ions towards the electrode surface be­ branched electrolytic deposition growth pattern is observed for 0.025 M
comes rate-limiting when the concentration of the ions (CCu2+ ,surface .) on electrolyte concentration compared to more uniform and slow growth in
the electrode surface falls to zero due to the high rate of the reaction. In the 0.1 M electrolyte concentration system. The model shows a high
that situation, the relative effect of electro-migration and diffusion of level of accuracy in predicting electrodeposition growth. The average
ions determines the morphology of the growing unit. If error for any selected parameter is found to be less than around 11
electro-migration predominates over diffusion, a sharp tip rises much percent, irrespective of the electrolytic concentration. Therefore, it is
longer forming dendrites instead of uniform deposition due to the high evident, after analyzing multiple matrix parameters, that the
field strength over that point. At a higher concentration of electrolyte ConvLSTM-based deep learning model is capable of predicting electro­
solution, it is already established that comparatively lower growth ve­ deposition growth with reasonable accuracy. All the results indicate that
locity is obtained due to enhanced diffusive flux and inter-ionic inter­ the deep learning approach can be used as a remarkable tool and will be
action [44] as discussed earlier. Therefore, the difference in morphology certainly useful for electrodeposition studies in near future.
between the two aforementioned concentrations (0.025 M and 0.1 M) is
quite reasonable. However, the present objective specifically deals with Author contributions
electrodeposition prediction by analyzing real-time captured images.
The detailed mechanism behind the formation of dendrites is not R.K. and T.D. worked on data curation and data validation. R.K.
required to foresee the growth. It is also observed that the model is helped in the data visualization and further investigation. Idea
predicting the electrodeposition growth more accurately at a later stage conceptualization, methodology development, and project administra­
of the experiment. Owing to the rapid growth rate and complex nature of tion, are done by H.H. and M.C. The model development and formal
the electrodeposition, the model lacks to predict an accurate outcome at analysis is performed by H.H. and R.K. Project supervision, funding, and
the early stage of the experiment. This shortcoming can be overcome by resource acquisition are managed by M.C. All authors contributed to
capturing more data points during the initial growth phase. The limi­ writing the original draft, editing and reviewing the manuscript.
tations of the high-speed image capturing capability set a restriction on
the frame rate of the captured images from the experiments. However, Declaration of Competing Interest
over time, the rate of electrodeposition growth decreases, and due to the
availability of sufficient data points for the same interval, the pre­ The authors declare that they have no known competing financial
dictions become significantly accurate . To further generalize the model, interests or personal relationships that could have appeared to influence
more experiments can be conducted by varying experimental conditions the work reported in this paper.
like the voltage, and inter-electrode gap, and the model can be trained
on a new set of images. This would lead to a more robust model. As the Data Availability
existing model is data-driven and has the capability of predicting an
accurate future electrodeposition growth, a similar concept can be used All the relevant have been uploaded with the manuscript.
for other experiments as well where a prediction of the future experi­
mental outcome is required. However, the model has to be trained
accordingly to get an accurate result. Acknowledgement

4. Conclusion The authors gratefully acknowledge the use of the High-Power

Computing facility “PARAM Shakti” at the Indian Institute of Technol­
A deep learning-based image processing method is proposed to ogy Kharagpur, India. One of the authors, Dr. M.C., gratefully ac­
determine the dendritic growth in rechargeable batteries. An knowledges the financial support provided by SRIC through ISIRD (No.

10
R. Kumar et al. Energy and AI 10 (2022) 100203

IIT/SRIC/2019-20/10) scheme, part of which was used to develop the [24] Fleury V, Rosso M, Chazalviel JN, Sapoval B. Experimental aspects of dense
morphology in copper electrodeposition. Phys Rev A 1991;44(10):6693–705.
other computational facilities used in this work
https://doi.org/10.1103/PhysRevA.44.6693.
[25] Wood KN, Kazyak E, Chadwick AF, Chen KH, Zhang JG, Thornton K, et al.
Supplementary materials Dendrites and pits: untangling the complex behavior of lithium metal anodes
through operando video microscopy. ACS Cent Sci 2016;2(11):790–801. https://
doi.org/10.1021/ACSCENTSCI.6B00260/SUPPL_FILE/OC6B00260_SI_001.MOV.
Supplementary material associated with this article can be found, in [26] Zhang B, Yang G, Li C, Huang K, Wu J, Hao S, et al. Electrochemical behaviors of
the online version, at doi:10.1016/j.egyai.2022.100203. hierarchical copper nano-dendrites in alkaline media. Nano Res 2018;11(8):
4225–31. https://doi.org/10.1007/s12274-018-2010-3.
[27] Nikolić ND, Popov KI, Pavlović LJ, Pavlović MG. M.orphologies of copper deposits
References obtained by the electrodeposition at high overpotentials. Surf Coatings Technol
2006;201(3–4):560–6. https://doi.org/10.1016/j.surfcoat.2005.12.004.
[1] Zhang LC, Chen LY, Wang L. Surface modification of titanium and titanium alloys: [28] Tan M, Harb JN. Additive behavior during copper electrodeposition in solutions
technologies, developments, and future interests. Adv Eng Mater 2020;22(5):1–37. containing Cl− , PEG, and SPS. J Electrochem Soc 2003;150(6):C420. https://doi.
https://doi.org/10.1002/adem.201901258. org/10.1149/1.1570412.
[2] Kalinina E, Pikalova E. Opportunities, challenges and prospects for [29] Mistry A, Franco AA, Cooper SJ, Roberts SA, Viswanathan V. How machine
electrodeposition of thin-film functional layers in solid oxide fuel cell technology. learning will revolutionize electrochemical sciences. ACS Energy Lett 2021;6(4):
Materials (Basel) 2021;14(19). https://doi.org/10.3390/ma14195584. 1422–31. https://doi.org/10.1021/ACSENERGYLETT.1C00194/SUPPL_FILE/
[3] Xu J, Feng W, Shan H, Jiang L, Tang L. Application of electrodeposition method in NZ1C00194_SI_001.PDF.
upgrading mortar durability. J Adv Concr Technol 2015;13(7):367–72. https://doi. [30] De Bézenac E, Pajot A, Gallinari P. Deep learning for physical processes:
org/10.3151/jact.13.367. incorporating prior scientific knowledge. In: 6th Int. Conf. Learn. Represent. ICLR
[4] Bard AJ, Faulkner LR. E.lectrochemical methods - Fundamentals and applications. 2018 - Conf. Track Proc; 2017. https://doi.org/10.48550/arxiv.1711.07970.
2nd. New York: WILEY; 2001. [31] Fischer P, Dosovitskiy A, Ilg E, Häusser P, Hazırbaş C, Golkov V, et al. FlowNet:
[5] Mladenović IO, Lamovec JS, Radović DGV, Vasilić R, Radojević VJ, Nikolić ND. M. learning optical flow with convolutional networks. 2015. https://doi.org/
orphology, structure and mechanical properties of copper coatings 10.48550/arxiv.1504.06852.
electrodeposited by pulsating current (PC) regime on Si(111). Metals (Basel) 2020; [32] Attia PM, Grover A, Jin N, Severson KA, Markov TM, Liao, et al. Closed-loop
10(4):1–21. https://doi.org/10.3390/met10040488. optimization of fast-charging protocols for batteries with machine learning. Nat
[6] Nikolić ND, Živković PM, Pavlović MG, Baščarević Z. Overpotential controls the 2020;578(7795):397–402. https://doi.org/10.1038/s41586-020-1994-5. 2020
morphology of electrolytically produced copper dendritic forms. J Serb Chem Soc 5787795.
2019;84(11):1209–20. https://doi.org/10.2298/JSC190522066N. [33] Ng MF, Zhao J, Yan Q, Conduit GJ, Seh ZW. Predicting the state of charge and
[7] Devos O, Gabrielli C, Beitone L, Mace C, Ostermann E, Perrot H. Growth of health of batteries using data-driven machine learning. Nat Mach Intell 2020;2(3):
electrolytic copper dendrites. I: current transients and optical observation. 161–70. https://doi.org/10.1038/s42256-020-0156-7.
J Electroanal Chem 2007;606(2):75–84. https://doi.org/10.1016/J. [34] Tu H, Moura S, Fang H. Integrating electrochemical modeling with machine
JELECHEM.2007.03.028. learning for lithium-ion batteries. 2021.
[8] Xu W, Wang J, Ding F, Chen X, Nasybulin E, Zhang Y, et al. Lithium metal anodes [35] Katirci R, Aktas H, Zontul M. The prediction of the ZnNi thickness and Ni% of ZnNi
for rechargeable batteries. Energy Environ Sci 2014;7(2):513–37. https://doi.org/ alloy electroplating using a machine learning method. Trans Inst Met Finish 2021;
10.1039/C3EE40795K. 99(3):162–8. https://doi.org/10.1080/00202967.2021.1898183.
[9] Zubi G, Dufo-López R, Carvalho M, Pasaoglu G. The lithium-ion battery: state of the [36] Sanchez AJ, Kazyak E, Chen Y, Chen KH, Pattison ER, Dasgupta NP. Plan-view
art and future perspectives. Renew Sustain Energy Rev 2018;89:292–308. https:// operando video microscopy of Li metal anodes: identifying the coupled
doi.org/10.1016/J.RSER.2018.03.002. relationships among nucleation, morphology, and reversibility. ACS Energy Lett
[10] Shah R, Mittal V, Matsil E, Rosenkranz A. Magnesium-ion batteries for electric 2020;5(3):994–1004. https://doi.org/10.1021/ACSENERGYLETT.0C00215/
vehicles: current trends and future perspectives. Adv Mech Eng 2021;13(3):1–9. SUPPL_FILE/NZ0C00215_SI_005.MP4.
https://doi.org/10.1177/16878140211003398. [37] Li Y, Zhang G, Sha L, Zhao W, Chen B, Wang D, et al. Phase-field simulation of
[11] Li L, Li S, Lu Y. Suppression of dendritic lithium growth in lithium metal-based dendrite growth in rechargeable batteries. Energy Storage Sci Technol 2022;11(3):
batteries. Chem Commun 2018;54(50):6648–61. https://doi.org/10.1039/ 929.
C8CC02280A. [38] Monroe C, Newman J. The impact of elastic deformation on deposition kinetics at
[12] Li Y, Qi Y. Energy landscape of the charge transfer reaction at the complex Li/SEI/ lithium/polymer interfaces. J Electrochem Soc 2005;152(2):A396. https://doi.org/
Electrolyte interface. Energy Environ Sci 2019;12(4):1286–95. https://doi.org/ 10.1149/1.1850854.
10.1039/c8ee03586e. [39] Fujiyoshi H, Hirakawa T, Yamashita T. Deep learning-based image recognition for
[13] Tewari D, Rangarajan SP, Balbuena PB, Barsukov Y, Mukherjee PP. M.esoscale autonomous driving. IATSS Res 2019;43(4):244–52. https://doi.org/10.1016/J.
anatomy of dead lithium formation. J Phys Chem C 2020;124(12):6502–11. IATSSR.2019.11.008.
https://doi.org/10.1021/acs.jpcc.9b11563. [40] Nishida T, Nishikawa K, Rosso M, Fukunaka Y. Optical observation of Li dendrite
[14] Chen KH, Wood KN, Kazyak E, Lepage WS, Davis AL, Sanchez AJ. Dead lithium: growth in ionic liquid. Electrochim Acta 2013;100:333–41. https://doi.org/
mass transport effects on voltage, capacity, and failure of lithium metal anodes. 10.1016/j.electacta.2012.12.131.
J Mater Chem A 2017;5(23):11671–81. https://doi.org/10.1039/C7TA00371D. [41] Bai P, Li J, Brushett FR, Bazant MZ. Transition of lithium growth mechanisms in
[15] Xu K. Nonaqueous liquid electrolytes for lithium-based rechargeable batteries. liquid electrolytes. Energy Environ Sci 2016;9(10):3221–9. https://doi.org/
Chem Rev 2004;104(10):4303–417. https://doi.org/10.1021/cr030203g. 10.1039/C6EE01674J.
[16] Wang S, Rafiz K, Liu J, Jin Y, Lin JYS. Effects of lithium dendrites on thermal [42] Vishnugopi BS, Hao F, Verma A, Mukherjee PP. Surface diffusion manifestation in
runaway and gassing of LiFePO4batteries. Sustain Energy Fuels 2020;4(5): electrodeposition of metal anodes. Phys Chem Chem Phys 2020;22(20):11286–95.
2342–51. https://doi.org/10.1039/d0se00027b. https://doi.org/10.1039/D0CP01352H.
[17] Schilardi P, Méndez S, Salvarezza RC, Arvia AJ. E.volution of the growth front for [43] Nie K, Hong Y, Qiu J, Li Q, Yu X, Li H, et al. Interfaces between cathode and
copper electrodeposition followed by real time imaging. Langmuir 1998;14(15): electrolyte in solid state lithium batteries: challenges and perspectives. Front.
4308–14. https://doi.org/10.1021/LA971358K. Chem. 2018;6(DEC):616. https://doi.org/10.3389/FCHEM.2018.00616/BIBTEX.
[18] Rodriguez R, Loeffler KE, Nathan SS, Sheavly JK, Dolocan A, Heller A, et al. In situ [44] Dhara T, Ghosh UU, Ghosh A, Vishnugopi BS, Mukherjee PP, DasGupta S.
optical imaging of sodium electrodeposition: effects of fluoroethylene carbonate. Mechanistic underpinnings of morphology transition in electrodeposition under
ACS Energy Lett 2017;2(9):2051–7. https://doi.org/10.1021/ the application of pulsatile potential. Langmuir 2022;38:4879–86. https://doi.org/
acsenergylett.7b00500. 10.1021/acs.langmuir.2c00066.
[19] Yufit V, Tariq F, Eastwood DS, Biton M, Wu B, Lee PD. Operando visualization and [45] Davidson R, Verma A, Santos D, Hao F, Fincher CD, Zhao D, et al. Mapping
multi-scale tomography studies of dendrite formation and dissolution in zinc mechanisms and growth regimes of magnesium electrodeposition at high current
batteries. Joule 2019;3(2):485–502. https://doi.org/10.1016/J. densities. Mater Horizons 2020;7(3):843–54. https://doi.org/10.1039/
JOULE.2018.11.002. c9mh01367a.
[20] Fang J, You H, Zhu C, Kong P, Shi M, Song X, et al. Thermodynamic and kinetic [46] Getreuer P. Chan-Vese segmentation. Image Process. Line 2012;2:214–24. https://
competition in silver dendrite growth. Chem Phys Lett 2007;439(1–3):204–8. doi.org/10.5201/IPOL.2012.G-CV.
https://doi.org/10.1016/J.CPLETT.2007.03.046. [47] Cho K, Van Merriënboer B, Gulcehre C, Bahdanau D, Bougares F, Schwenk H, et al.
[21] Davidson R, Verma A, Santos D, Hao F, Fincher C, Xiang S, et al. Formation of Learning phrase representations using RNN encoder-decoder for statistical machine
magnesium dendrites during electrodeposition. ACS Energy Lett 2019;4(2):375–6. translation. In: EMNLP 2014 - 2014 Conf. Empir. Methods Nat. Lang. Process. Proc.
https://doi.org/10.1021/ACSENERGYLETT.8B02470/SUPPL_FILE/NZ8B02470_ Conf; 2014. p. 1724–34. https://doi.org/10.48550/arxiv.1406.1078.
SI_004.AVI. [48] Graves A, Mohamed AR, Hinton G. Speech recognition with deep recurrent neural
[22] Edwards SF, Wilkinson DR. The surface statistics of a granular Aggregate. In: Proc. networks. In: ICASSP, IEEE Int. Conf. Acoust. Speech Signal Process. - Proc; 2013.
R. Soc. London. A. Math. Phys. Sci. 381; 1982. p. 17–31. https://doi.org/10.1098/ p. 6645–9. https://doi.org/10.48550/arxiv.1303.5778.
rspa.1982.0056. [49] Hewamalage H, Bergmeir C, Bandara K. Recurrent neural networks for time series
[23] Schneider NM, Park JH, Grogan JM, Steingart DA, Bau HH, Ross FM. Nanoscale forecasting: current status and future directions. Int J Forecast 2021;37(1):
evolution of interface morphology during electrodeposition. Nat Commun 2017;8 388–427. https://doi.org/10.1016/J.IJFORECAST.2020.06.008.
(1):1–10. https://doi.org/10.1038/s41467-017-02364-9. 2017 81.

11
R. Kumar et al. Energy and AI 10 (2022) 100203

[50] Pollastri G, Przybylski D, Rost B, Baldi P. Improving the prediction of protein activity towards the hydrogen evolution reaction. ChemSusChem 2018;11(2):
secondary structure in three and eight classes using recurrent neural networks and 367–75. https://doi.org/10.1002/cssc.201701938.
profiles. Proteins 2002;47(2):228–35. https://doi.org/10.1002/PROT.10082. [60] Saab R, Sultan RDensity. Fractal angle, and fractal dimension in linear zn
[51] Shi X, Chen Z, Wang H, Yeung DY, Wong WK, Woo WC. C.onvolutional LSTM electrodeposition morphology. J Non-Equilibrium Thermodyn 2005;30(4):321–36.
network: a machine learning approach for precipitation nowcasting. Adv Neural https://doi.org/10.1515/JNETDY.2005.023.
Inf Process Syst 2015;2015-Janua:802–10. https://doi.org/10.48550/ [61] Selis LA, Seminario JM. D.endrite formation in Li-metal anodes: an atomistic
arxiv.1506.04214. molecular dynamics study. RSC Adv 2019;9(48):27835–48. https://doi.org/
[52] Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by 10.1039/C9RA05067A.
reducing internal covariate shift. In: 32nd Int. Conf. Mach. Learn. ICML 2015. 1; [62] OpenCV: findcontour and contour approximation method. 2022. https://docs.open
2015. p. 448–56. https://doi.org/10.48550/arxiv.1502.03167. cv.org/4.x/d4/d73/tutorial_py_contours_begin.html.
[53] Islam MZ, Islam MM, Asraf A. A combined deep CNN-LSTM network for the [63] OpenCV: contour features. 2022. https://docs.opencv.org/4.x/dd/d49/tuto
detection of novel coronavirus (COVID-19) using X-ray images. Informatics Med rial_py_contour_features.html.
Unlocked 2020;20:100412. https://doi.org/10.1016/J.IMU.2020.100412. [64] Issa MA, Issa MA, Islam MS, Chudnovsky A. Fractal dimension-a measure of
[54] Tran D, Bourdev L, Fergus R, Torresani L, Paluri M. Learning spatiotemporal fracture roughness and toughness of concrete. Eng Fract Mech 2003;70(1):125–37.
features with 3D convolutional networks. In: Proc. IEEE Int. Conf. Comput. Vis. https://doi.org/10.1016/S0013-7944(02)00019-X.
2015 Inter; 2015. p. 4489–97. https://doi.org/10.1109/ICCV.2015.510. [65] Zhihui N, Lichun W, Ming-Hui W, Jing Y, Qiang Z. The fractal dimension of river
[55] Zeiler MD. A.dadelta: an adaptive learning rate method. 2012. https://doi.org/ length based on the observed data. J Appl Math 2013;2013. https://doi.org/
10.48550/arxiv.1212.5701. 10.1155/2013/327297.
[56] Asamoah D, Ofori E, Opoku S, Danso J. Measuring the performance of image [66] Richardson LF. The problem of contiguity: an appendix to statistics of deadly
contrast enhancement technique. Int J Comput Appl 2018;181(22):6–13. https:// quarrels. Gen Syst Yearb 1961;6:139–87.
doi.org/10.5120/IJCA2018917899. [67] Block A, Von Bloh W, Schellnhuber HJ. Efficient box-counting determination of
[57] Sara U, Akter M, Uddin MS, Sara U, Akter M, Uddin MS. I.mage quality assessment generalized fractal dimensions. Phys Rev A 1990;42(4):1869. https://doi.org/
through FSIM, SSIM, MSE and PSNR—a comparative study. J Comput Commun 10.1103/PhysRevA.42.1869.
2019;7(3):8–18. https://doi.org/10.4236/JCC.2019.73002. [68] Li J, Du Q, Sun C. An improved box-counting method for image fractal dimension
[58] skimage — skimage v0.19.2 docs. 2022. https://scikit-image.org/docs/stable/api estimation. Pattern Recognit 2009;42(11):2460–9. https://doi.org/10.1016/J.
/skimage.html. PATCOG.2009.03.001.
[59] Isarain-Chávez E, Baró MD, Alcantara C, Pané S, Sort J, Pellicer E. Micelle-assisted [69] Mistry A, Srinivasan V. On our limited understanding of electrodeposition. MRS
electrodeposition of mesoporous Fe–Pt smooth thin films and their electrocatalytic Adv 2019;4(51):2843–61. https://doi.org/10.1557/adv.2019.443.

12