Vol. 06, No. 1 (2024) 271-280, doi: 10.24874/PES.SI.24.02.

010

Proceedings on Engineering
Sciences
www.pesjournal.net

PREDICTING ALUMINA COMPOSITES’

MECHANICAL CHARACTERISTICS USING A
MACHINE LEARNING APPROACH

Ashwini Kumar1
Arunkumar Devalapura Thimmappa Received 14.11.2023.
Received in revised form 27.12.2023.
Ritesh Kumar Accepted 05.01.2024.
Manali Gupta UDC – 004.85

ABSTRACT
Keywords:
Aluminium Alloys, Mechanical Obtaining the requisite properties in alloys is crucial problem in the
Characteristics, Machine Learning, production of aluminium components, requiring great deal of time and effort
Bayesian-fine tuned Adaptive Gated for investigation and experimentation. In this study, machine-learning
Recurrent Unit (B-AGRU), Tensile technique utilizing Bayesian-fine tuned Adaptive Gated Recurrent Unit (B-
Strength, Hardness. AGRU) to forecast the mechanical characteristics of aluminium alloys is
presented. Training and testing are conducted on dataset, which has
undergone comprehensive preparation process that includes cleaning and Z-
score normalization. Principal Component Analysis (PCA) is used for feature
extraction to increase algorithmic efficiency. The GRU approach, which is
implemented in Python, hardness and yield strength, leading in more accurate
findings. When compared to standard methodologies, process saves
significant time and energy, as evidenced by metrics such as RMSE-20%,
MAE-10% and R-squared-97%. This study reveals B-AGRU-based machine
learning as a feasible strategy for enhancing efficiency and sustainability in
forecasting mechanical properties of aluminium alloys, paving the way for
wider application in industrial sector.
© 2024 Published by Faculty of Engineeringg

1. INTRODUCTION alumina composite mechanical characteristics is crucial

for optimizing their performance in a wide range of
In the field of sophisticated materials, alumina applications, from spacecraft structural components to
composites stand out as a collection of substances with cutting-edge medical implants (Boopalakrishnan et al.,
excellent mechanical properties, making them useful in (2023)).
industries ranging from aerospace to biomedical
engineering (Raju et al., (2020)). Because of their The backbone of these composites is alumina, which is
unique combination of strength, hardness and thermal known for its hardness and wear resistance. It is the
stability, these aluminum oxide (Al2O3) composites are deliberate incorporation of other elements, typically in
suitable for a wide range of applications (Agrawal and the form of reinforcing phases or additives that unlocks
Satapathy (2019)). Understanding and anticipating the full potential of alumina composites. Secondary

1
Corresponding author: Ashwini Kumar
Email: [email protected] 271
 Kumar et al., Predicting alumina composites’ mechanical characteristics using a machine learning approach

phases can contain ceramics, polymers, or even metallic Najjar et al., (2022) used a micromechanics model and
components, each of which provides unique properties finite element simulation to estimate elastic modulus
to the resulting composite material (Kota et al., (2022)). in Cu-Al2O3 nanocomposites utilizing a rule of
The interaction of the main alumina matrix with various mixing and grain evolution. Using “finite element
additives yields a material with customized properties, (FE)” outputs and experimental data, a Random vector
allowing engineers to construct materials with specified functional link network machine learning model was
performance features (Lakhdar et al., (2021)). developed, attaining a high R¬2 (0.99) for predicting
yield strength in micro indentation tests on produced
Alumina composites' physical properties are multifaceted, nanocomposites. The technique effectively combines
encompassing a range of characteristics including tensile modeling, simulation & machine learning to forecast
strength, hardness, fracture toughness and fatigue mechanical parameters with excellent agreement with
resistance. To learn the possibilities of these materials, data from experiments.
researchers and engineers use predictive models that
account for a variety of factors influencing the physical Yu et al., (2021) predicted the mechanical
properties of the composites (Goswami et al., (2022)). characteristics of aluminum alloys using a “deep
There is a complex interplay between the composition, neural network (DNN)” that was pre-trained for
processing methods and microstructure of alumina parameter initialization. The results showed that DNN
composites, as these factors have a significant impact on is more accurate and more broadly applicable than
their mechanical responses (Cygan et al., (2019)). “support vector regression (SVR) and shallow neural
networks (SNN)”, indicating its potential for a data-
Predicting mechanical properties is essential for driven material design that can be extended to other
guaranteeing the dependability and safety of materials. However, DNN's limitations include its
components composed of alumina composites along reliance on small datasets.
with optimizing material efficiency (Sharma et al.,
(2022)). In mission-critical situations, accurate VarolÖzkavak et al., (2023) used artificial intelligence
prediction of the mechanical characteristics of methods “convolutional neural network, deep learning
aluminium composite might be the distinction between algorithm, artificial neural network and random forest
success and disaster, such as in aircraft applications regression (CNN, ANN, RFR)” to be prepared the
where weight, strength and thermal stability are mechanical properties (hardness, bending strength) of
essential elements (Altıparmak et al., (2022)). AA 2024 Al alloys following being aged at different
Analogously, in the biomedical domain, predicting the temps and duration. The best results were obtained for
mechanical characteristics of alumina composites is PM and FD 2024 Al alloys with CNN “(RMSE
critical to the design of implants that can endure the 0.09068, R-Squared 0.93476, MAE 0.06734)” as well
physiological stresses imposed by the human body as “(RMSE 0.08578, R-Squared 0.94166, MAE
(Palmero (2019)). 0.06212)”, respectively limitations included reliance
on available experimental data.
Banerjee et al., (2020) utilized an artificial intelligence-
driven approach, utilizing artificial neural networks Arpitha et al., (2023) used a central composites design-
(ANN) and genetic algorithms, to optimize the response surface technique & artificial neural network
composition size; based on conflicting goal functions architecture to optimize the synergistic impact of
generated from literature data, they investigated the sugarcane bagasse & aluminum micro-fillers (1.38 to
mechanical and morphological features of alumina- 5.62 wt. %) of epoxy polymer composites. Both
reinforced composites made of aluminum matrix for sugarcane bagasse and alumina micro-fillers affect
better tribo-mechanical properties. The technique mechanical properties, with 3.5 wt. % optimizing
produces Pareto solutions for customized composite flexural qualities. Epoxy resin composites have better
design, with findings proving improved performance physical, thermal and mechanical characteristics, yet
while accepting study constraints. microscopy shows uniform filler distribution, verifying
the model's correctness.
Liu et al., (2022) predicted the mechanical
characteristics of graphene-reinforced aluminum Kordijazi et al., (2021) evaluated the wetting properties
nanocomposites by combining molecular dynamics of graphite, NiAl3 and SiC-containing Al-Si alloys and
simulations with machine learning (ML) approaches. Al composites with metal matrix using theoretical,
Incorporating graphene volume fractions, alignment experimental & machine learning (ML) approaches. ML
angle, chirality along with the temperature into MD models with significant positive correlations (r > 0.9)
simulations results in the development of ML models between predicted and observed water contact angles,
for Young's modulus and ultimate tensile strength, as showing robustness, include regression, ANN, “chi-
well as the modification of the Halpin-Tsai model for square automatic interaction detection (CHAID)”,
enhanced Young's modulus accuracy in forecasting, XGBoost and random forest. Mechanical abrasion,
allow for more practical assessment and development of etching and graphite addition all raise contact angles.
graphene-aluminum nanocomposite material structures. With constraints due to the complexity of multiphase

272
 Proceedings on Engineering Sciences, Vol. 06, No. 1 (2024) 271-280, doi: 10.24874/PES.SI.24.02.010

alloy systems, they give interesting insights for insights into influencing features for steel creep
predicting wetting qualities and comprehending properties. Limitations include the dependence on
physical events connected with alloy wettability. conventional creep research knowledge and potential
generalization issues.
Rajput et al., (2022) aims to enhance stir casting for
marine and aerospace applications of hybrid metallic Liu et al., (2023) suggested a deep learning-based
matrix composites (Al-HMMC) based on aluminum. technique based on convolutional neural networks to
They use a decision tree regression model to forecast forecast damage and deformation in self-piercing
Al-HMMC's mechanical properties properly. The model riveted joints of aluminum alloys and carbon fiber-
predicts Ultimate Tensile Strength with 92.029% reinforced composites. The model solves the problems
accuracy. In experimental validation on Al7075 and of complicated simulation and high experimental costs
Al6061 matrices, anticipated and real UTS values differ in self-piercing riveted (SPR) joint analysis by reliably
by less than 10%. Machine learning can minimize predicting section parameters with average accuracies of
experimental costs and time for improving Al-HMMC 95.80% for riveted head height, 95.68% for residual
characteristics, according to the research. thickness and 92.40% for rivet spread.

According to Deng et al., (2020), machine learning Hajilounezhad et al., (2021) effort was to accurately
(ML), namely the “SMOreg/puk technique”, can classify and predict the morphology and mechanical
forecast the Cu-Al alloy's mechanical characteristics in properties of vertically oriented carbon nanotube (CNT)
metallic powder with an accuracy that is higher than forests through the use of an image-based machine
that of five other models. “Cu-12Al6-Ni” alloy with learning technique and a physics-based simulation
expected “tensile force (390 MPa) & hardness (HB called CNT Net. This will enable high-throughput
139)” was produced utilizing the SMOreg/puk model, material discovery. By using synthetic images to train
offering a rapid and efficient method for material CNT Net, the technique has shown >91% accuracy in
synthesis & characterization using chemical categorizing CNT characteristics and superiority in
composition and porosity as descriptors. The research stiffness and buckling load prediction over physical
provides important insights into composition design, parameter-based predictors. Constraints include the
even though model generalization can have limitations. intricacy of the regulating processes and the extensive
range of experimental parameters.
According to Devadiga et al., (2019), "scanning electron
microscope (SEM)" blend powder morphology as well Katırcı and Yıldız (2023) used ML algorithms Multi-
as sintered composite structure characterization are used layer Perceptron, Random Forest and Extreme Gradient
to predict density and hardness in "multi-walled carbon Boosting (MLP, RF, XGBoost) with improved hyper
nanotubes (MWCNT)" coupled with "fly ashes parameters to forecast Al2O3-Cr2O3 ceramic fracture
(FA/Al)" composites using artificial neural networks behavior. RF predicted Cr2O3 ratio as critical for
(ANN) that connect them with encouragement quantity, relative density, diameter for fracture strength and
ball milling duration and sintering time exploratory thickness for total crack length, matching actual results.
mechanical property determination. Based on how Genetic Algorithm discovered best solution (0.7%
property dispersion & augmentation are affected by Cr2O3, 28.5 mm diameter, 2.2 mm thickness, 325.8
reinforcing content, ball milling duration and sintering MPa fracture strength). Surface response
time, ANN predicts mechanical properties. The trial's experimentation and leave-one-out cross-validation
scope can restrict the range of variables and their were used. Results help forecast material properties, but
generalizability. generalization concerns and dataset features restrict
them.
Wiciak-Pikuła et al., (2020) used “multilayered
perceptron (MLP)” artificial neural networks to Thirumoorthy et al., (2019) used of stir casting to
develop effective face grinding tool usage models for produce Al6061 MMCs with blended MgO and Si3N4,
forecasting with “aluminum matrix composites with tensile properties assessed using a novel hybrid
(AMC)” that include 10% SiC. The models were approach combining “K-nearest neighbor (KNN)” and
evaluated using mean square error and limitations in “ant lion optimization (ALO)” methods. Validated
“tool flank wear (〖VB〗_B)” as well as “tool corner against a decision tree (DT) classifier, the results show
wear (〖VB〗_C)” forecasting during machining were that KNN-ALO was effective at predicting tensile and
addressed. hardness properties of composites, contributing to
advancements in material processing and
Wang et al., (2021) used machine learning to predict characterization.
the creep life of Cr-Mo steel using time-temperature
parameters Larson-Miller parameter, Manson-Haferd The B-AGRU model's development overcame
parameter and Manson-Succop parameter (LMP, challenges in capturing complicated patterns and
MHP, MSP) demonstrating enhanced precision with connections, hence addressing the difficulty of
random forest algorithms while offering valuable forecasting mechanical features in Alumina Composites.

273
 Kumar et al., Predicting alumina composites’ mechanical characteristics using a machine learning approach

1.1 Contribution technique. Gathering data set, the data cleaning and
z-score normalization is used for pre-processing. The
feature extraction using a PCA, Figure 1 illustrates
• The study gathers dataset for testing and training. the structure of methodology.
• Pre-processing data entails Z-score normalization
and robust cleaning for noise reduction, which
guarantee fair along with consistent comparisons
and improve dataset quality.
• Using PCA is a critical step that reduces the
dimensionality of the data while preserving
important information, improving algorithmic
efficiency.
• The B-AGRU method is a powerful machine-
learning approach used in this study to forecast the Figure 1. Overview of methodology.
mechanical characteristics of aluminum alloys.
2.1 Data set
The remaining part of the study: part 2 discusses the
methodology, part 3 assesses the efficiency of the In this research, a total of 130 datasets were acquired
proposed method and part 4 concludes the paper. from material matching experiments (Devi et al.,
(2020)). These datasets encompass various
2. METHODOLOGY compositions of aluminum alloys, yield strength,
hardness and tensile strength of aluminum composites,
In the section, the research forecasted the mechanical with matching values. The specific details of these
properties of aluminum alloys using the potent values are presented in Table 1.
machine-learning technique known as the B-AGRU

Table 1. Mechanical properties of aluminum alloys.

Sl. Cr Mn Si Cu Fe Z Al Zn Mg Yield Tensile
Other Hardness
number % % % % % r% % % % strength strength
1 0 0.3 0 79.1 4.7 0 11 0 0 0 850 185 -70
2 0 0 0.6 0 0.6 0.1 96.3 0 1.5 0 135 60 -44
3 0 1.2 2.2 0.2 0.7 0 95 0 0.5 0 200 170 -44
- - - - - - - - - - - - - -
- - - - - - - - - - - - - -
128 0 0.55 11 0.35 0.65 0 86.05 0.35 0.45 0.2 180 380 -80
129 0 0.45 1.5 0.05 0.55 0 90.5 0.1 6.5 0.2 219 250 -65
130 0.1 1 0.8 4.6 0.8 0 87.8 0.8 1.8 0.2 370 250 -100

2.2 Data pre-processing quality. As part of this procedure, missing values are
handled via imputation or elimination; errors in data inputs
Data pre-processing is the process of preparing and are identified and corrected as well as formats are
altering raw data before utilizing it in machine learning standardized to promote uniformity. To verify the integrity
or analytical algorithms. To provide a constant scale for of the data, it entails evaluating it against predetermined
accurate modeling and analysis throughout the rules or statistical criteria. Data cleaning seeks to improve
prediction phase, Z-score normalization is utilized to the accuracy and efficacy of data-driven decision-making
normalize the features. Data cleaning is used to resolve processes by methodically resolving these problems and
missing or incorrect values as part of the data pre- producing a cleaned dataset that provides a strong basis for
treatment procedure to forecast the mechanical further analysis and modeling.
characteristics of Alumina Composites.
2.2.2 Z-score normalization
2.2.1 Data cleaning
In predicting the mechanical properties of Alumina
Data cleaning, an essential part of data preparation, is Composites, Z-score normalization entails adjusting the
locating and fixing mistakes, inconsistencies and raw data to have a mean of zero and a standard deviation
inaccuracies in a dataset to improve its dependability and of one, maintaining uniform scales for different features

274
 Proceedings on Engineering Sciences, Vol. 06, No. 1 (2024) 271-280, doi: 10.24874/PES.SI.24.02.010

and improving model performance. A data preprocessing The orthonormal vectors that constitute the S-matrix of
method called Z-score normalization, commonly called values are obtained through a linear mixture of
standardization, was used to convert numerical data to components from the W-matrix. This combination
comply with a standard distribution. It includes dividing describes the relationship between the samples. The
by the standard deviation after considering the data's 𝑆matrix of covariance is defined as Eq. (4):
mean. The values of attribute B are standardized under
their standard deviation and mean using the 𝐷𝑆 = 𝑌𝐷𝑤 𝑌𝑆 (4)
normalization approach. The following Eq. (1) converts
̂:
a value of B to𝑏 The variable 𝐷𝑤 represents the matrix of covariance for
the variable𝑊.
𝑏−𝜇(𝑏)
̂=
𝑏 (1)
𝜕(𝑏)
The weighting matrices 𝑌 can be obtained by solving
the eigenvalue Eq. (5):
Where 𝜕(𝑏) reflects the properties′ standard deviation
𝐵, &𝜇(𝑏) shows the average value.
The method works well in stationary environments
(𝐷𝑠 − 𝜆𝐽)𝑓𝑗 = 0 (5)
because B's minimum and maximum values are known.
It struggles with non-stationary time series since their The covariance matrix contains the bilateral co-variances
standard deviation and mean change with time. between the several input variables. Subsequently, the
eigenvectors and eigenvalues for the matrix of covariance
2.3 Feature extraction using Principal are decomposed (as shown in Eq. (5). The resulting
component analysis (PCA) eigenvectors represent the new orthogonal elements,
referred to as “principal components, with their magnitude
To extract features and forecast the mechanical determined by the related Eigen values”. After arranging
properties of alumina composites, it was done using the eigenvalues and the related eigenvectors in a decreasing
principal component analysis (PCA). By lowering the order, the PC will follow the same order. The first principal
dimensionality of the data while maintaining crucial component will possess the highest variance, indicating the
information, this method aids in enhancing our most significant information. The subsequent principal
comprehension of the mechanical characteristics of the component will exhibit the second highest variance and so
material. PCA is a statistical method that uses data to on. It is important to mention that the main components
create a model. It takes a group of variables that are obtained are not associated with one another, regardless of
related to each other and transforms them into a least the correlation between the input parameters.
number of innovative parameters that are non-related to
each other. These innovative parameters retain a 2.4 Predicting mechanical properties of AL
significant amount of the actual data's information. Let alloy using for Bayesian- fine-tuned
𝑊 is the input data, where all column represents a Adaptive Gated Recurrent Unit (B-AGRU)
sequence of m-dimensional inputs. Furthermore, it is
important to note that the mean of each function in the The Gated Recurrent Unit (GRU) neural network and
set of values is zero (𝐸(𝑊) = 0). A data matrix in its Bayesian Optimization (BO) are used in the B-AGRU
original form consists of 𝑚 samples and 𝑛 variables, technique to forecast the mechanical characteristics of
which can be represented as following Eq. (2): aluminium alloys. BO is used to tune hyper parameters.
Algorithm 1 illustrate the B-AGRU
𝑤11 ⋯ 𝑤1𝑛
𝑊 = [𝑤1 , 𝑤2 , … , 𝑤𝑚 ]𝑆 = ( ⋮ ⋱ ⋮ ) (2) 2.4.1 Bayesian Optimization (BO)
𝑤𝑚1 ⋯ 𝑤𝑚𝑛
The Bayesian Optimization (BO) technique, utilized as a
Data on environmental parameters and performance model-based hyper parameter-tuning strategy, leverages
criteria can be converted into a new occurrence space surrogate function for simulating the conditional
using Principal Component Analysis (PCA), which probability of validation set performance with provided
preserves as much of the original data as is practical. To hyper parameters. Unlike grid / random searches, BO
do this, the directions with the most variance in the records earlier assessments, minimizing unnecessary
input information sets are found and they are predicted calculations for unfavorable hyper parameters. The
into a new subspace with the same or fewer extent than acquisition function chooses promising hyper parameters
the original space. for the next iteration, enhancing tuning effectiveness in a
shorter assessment time. The proposed model incorporates
Therefore, an orthonormal transformation 𝑌 can be BO algorithm techniques into the dynamic ensemble
employed to transfer𝑊 to a novel space 𝑆 in the module to achieve optimum hyper parameter tuning. The
following Eq. (3): BO approach is made up of five primary components:
hyper parameter distance, OF (forecasting error based
𝑆 = 𝑌𝑊 (3) enhanced validation data), acquisitions function, history of

275
 Kumar et al., Predicting alumina composites’ mechanical characteristics using a machine learning approach

assessments and surrogate function. The approach Where, a function called 𝑓ℎ𝑦𝑏𝑟𝑖𝑑 combines the Gated
leverages "tree-based Parzen window estimation (TPE)"
Recurrent Unit (GRU) and the outputs of Bayesian
for the probabilistic estimation of the surrogate function
Optimization (BO). The goal is to increase forecast
and the anticipated improvement acts as the acquisition
accuracy by utilizing the advantages of both methods.
function A, as shown in Eq. (6).
The GRU model's features or hyper parameters might be
ℊ∗
optimized using the BO as one potential strategy.
𝑎ℊ∗ (𝑉) = ∫−∞(ℊ∗ − ℊ)𝑄( ℊ|𝑣)𝑑ℊ, (6) Whereas the GRU detects sequential relationships in the
data, the BO model learns the mapping from input
Where𝑔 is the 𝜚ℱ and g is the 𝜚ℱ threshold, assuming parameters to desired attributes.
the hyper parameter selection𝑣.
The hybrid Eq. (14) is shown in the following simplified
2.4.2 Fine tuned Gated Recurrent Unit (GRU) form:

𝑦 = 𝐺𝑅𝑈(𝑋; 𝜃𝐺𝑅𝑈 ) + 𝐵𝑂(𝑋; 𝜃𝐵𝑂 ) (14)

The GRU is a “recurrent neural network (RNN)” that is an
expanded version of “long short-term memory (LSTM)”
unit. The GRU is a simplified model with two gate • The GRU model's output, represented
functions: the update gate, which determines how much as𝐺𝑅𝑈(𝑋; 𝜃𝐺𝑅𝑈 ), has parameters𝜃𝐺𝑅𝑈 .
previous information is preserved and the reset gate, which • The𝐵𝑂 model's output, represented as𝐵𝑂(𝑋; 𝜃𝐵𝑂 ),
commands the integration of past information with the has parameters𝜃𝐵𝑂 .
current input. This is in contrast to the LSTM's Input gate,
forgetting gate, & output gate are the three gate functions. It is possible to tune the𝜃𝐺𝑅𝑈 and 𝜃𝐵𝑂 model parameters
during training. With the assistance of the Bayesian
The formula for calculation at time 𝑡 is as following Eq. technique, this hybrid strategy aims to represent both
(7-10). the sequential dependencies in the data and the
optimization of the parameters. The specifics of our data
𝑟𝑡 = 𝜎𝑠𝑖𝑔 (𝑊𝑟 𝑥𝑡 + 𝑈𝑟 ℎ𝑡−1 ) (7) and the objectives of the prediction model will
determine how the implementation is performed out.
𝑧𝑡 = 𝜎𝑠𝑖𝑔 (𝑊𝑧 𝑥𝑡 + 𝑈𝑧 ℎ𝑡−1 ) (8)
Algorithm 1: B-AGRU
̃ = ∅𝑡𝑎𝑛ℎ (𝑧𝑡 = 𝜎𝑠𝑖𝑔 (𝑊ℎ 𝑥𝑡 + 𝑟𝑡 °𝑈ℎ ℎ𝑡−1 )
ℎ (9) Import necessary libraries
from agru model import create agrumodel # Assume
you have a function to create AGRU model
̃𝑡
ℎ𝑡 = (1 − 𝑧𝑡 )°ℎ𝑡−1 + 𝑧𝑡 °ℎ (10) from Bayesian optimization import optimize with BO
Assume you have a function for Bayesian Optimization
The current hidden node's candidate value is represented Sample data (replace with your actual data)
̃𝑡 in the formula above, while the activating value of
by ℎ X train, y train = generate training data()
the hidden node's output is represented byℎ𝑡 . The reset Step 1: Bayesian Optimization to optimize AGRU
gate is represented by𝑟𝑡 , while the update gate is by𝑧𝑡 . hyperparameters
°Denotes the multiplier element-wise. The activation Best params = optimize with bo(X train, y train)
Step 2: Train AGRU model with optimized
functions 𝜎𝑠𝑖𝑔 𝜑𝑡𝑎𝑛ℎ are responsible for activating control
hyperparameters
gates and candidate states, respectively. The 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 Agru model = create agrumodel(best params)
and 𝑡𝑎𝑛ℎ functions' expressions are Eq. (11-12): agru model. Fit(X train, y train)
Step 3: Make predictions using the trained AGRU
𝑠𝑖𝑔(𝑥) = (1 + 𝑒 𝑥 )−1 (11) model
Agru predictions = agru model. Predict(X test)
tanh(𝑥) = 2∗ 𝑠𝑖𝑔(𝑥) − 1 (12) Step 4: Use Bayesian Optimization to further refine
predictions
This study employs an enhanced GRU model to Refined params = optimize with bo(X test, agru
enhance the precision of wind power predictions, predictions)
aiming to achieve superior prediction accuracy. Step 5: Train final AGRU model with refined
Algorithm 1 illustrates the GRU. hyperparameters
Final agru model = create agrumodel(refined params)
The input parameters are represented by X and the final agru model. Fit(X train, y train)
analyses of alumina alloys mechanical properties were Step 6: Make final predictions
determined by Y. The following represents a Final predictions = final agru model. Predict(X test)
representation of the hybrid Eq. (13): Evaluate the final predictions
evaluate(final predictions, true values)
𝑦 = 𝑓ℎ𝑦𝑏𝑟𝑖𝑑 (𝑥) (13)

276
 Proceedings on Engineering Sciences, Vol. 06, No. 1 (2024) 271-280, doi: 10.24874/PES.SI.24.02.010

3. PERFORMANCE EVALUATION AND

DISCUSSION

The mechanical properties of Alumina composites are

predicted by implementing a B-AGRU in Python.
Key material attributes are included in the dataset.
Tensor Flow or PyTorch are used in the Python code
to build GRU. An adequate Graphics Processing Unit
(GPU) and a minimum of 8GB RAM is needed for
the experiment to educate the computer effectively.
We analyze the proposed method metrics of
performance: RMSE “(Root Mean Squared Error)”,
MAE “(Mean Absolute Error)” and R Squared.
Comparisons of our proposed method are deep neural
network (DNN) (Yu et al., 2021), support vector
regression (SVR) (Yu et al., 2021) and shallow neural Figure 3. B-AGRU-based Actual vs. Predicted Yield
network (SNN) (Yu et al., 2021). Strength.
Figure 2 shows a comparison of the B-AGRU The “actual and predicted” hardness values for B-
algorithms predicted and real tensile strength values. AGRU are shown in Figure 4. Notably, five or so data
The figure shows that the B-AGRU algorithm was points depart from the predicted fitting zone. Of them,
able to produce a better fit between the points and three or out of the total of 13 points do not match the
the regression line. Just one of the 13 data points, regression line, making them stand out. The outliers
which differ dramatically from the regression line, display a notable divergence from the regression line,
stands out as an outlier. The general data point indicating a noteworthy departure from the overall
alignment with the regression line highlights how pattern seen in the other data points.
well the algorithm predicts the tensile strength of
aluminum alloys.

Figure 4. B-GRU-based Actual vs. Predicted

Hardness value.
Figure 2. B-AGRU-based Actual vs. Predicted tensile
strength. A statistic known as the RMSEestimates the mean
variance between the results that were expected and
An analysis of yield strength predictions provided by those that weren't. the mean difference between results
the B-AGRU algorithm is shown in Figure 3 as predicted and as observed. The value of square root
about actual values. Closer examination of the plot of the average for the squared differences between
reveals that the residuals generated by the B-AGRU expected and actual values is calculated. A lower RMSE
method better fit the data. To be more precise, the suggests that the method B-AGRU predictions are more
figure demonstrates that the B-AGRU method fits accurate. Figure 5 and Table 2 illustrate the RMSE
the regression line better when evaluating the yield values. Compared to existing DNN-22, SNN-25 and
strength, which suggests a higher correlation SVR-28, our proposed method B-AGRU-20 was lower.
between the expected and actual values. Compared to existing approaches, the B-AGRU
improved the predicting alumina composites'
mechanical characteristics.

277
 Kumar et al., Predicting alumina composites’ mechanical characteristics using a machine learning approach

Table 3. Values of MAE.

Method MAE
DNN 13
SNN 15
SVR 14
B-AGRU [Proposed] 10

R-squared reflects the fraction of the variation in the

dependent variable that is predicted by the independent
variables. A higher R-squared value indicates a more
accurate match. R-squared is computed as the explained
variance divided by the total variance. Our proposed
method is higher B-AGRU- 97 than the existing
methods DNN-96, SNN-94 and SVR-93 as shown in
Figure 5. Outcome of RMSE. Figure 7 and Table 4. It shows that our proposed
method B-AGRU successfully predicts alumina
Table 2. Values of RMSE. composites' mechanical characteristics.

Method RMSE
DNN 22
SNN 25
SVR 28
B-AGRU [Proposed] 20

MAE is yet another statistic used to assess the

prediction model's accuracy. The percentage
differences between the expected and actual values
are averaged. Like RMSE, MAE evaluates the extent
to which the B-AGRU predictions correspond to the
actual values. It is frequently used in conjunction
with other assessment metrics to evaluate the B-
AGRU overall performance. Figure 6 and Table 3 Figure 7. Outcome the R2.
depict the response time values. Our suggested
approach was lesser than the existing method such as Table 4. Values of R2.
DNN-13, SNN-15 and SVR 14. The B-AGRU-10
enhanced in predicting alumina composites' Method 𝐑𝟐
mechanical characteristics when compared to existing DNN 96
methods. SNN 94
SVR 93
B-AGRU [Proposed] 97

4. CONCLUSION

The article offers a novel strategy for overcoming the

difficult task of acquiring the properties of aluminium
alloys required for component production. The study
presents a very successful prediction model for
mechanical properties including tensile strength,
hardness and yield strength by using machine learning,
most especially B-AGRU. The B-AGRU technique
performs better at predictions because it uses data for
training and testing, along with a careful preparation
procedure that includes cleaning and Z-score
normalization. The effectiveness of the technique is
Figure 6. Outcome of MAE. further improved by including Principal Component
Analysis (PCA) for feature extraction. Metrics such as

278
 Proceedings on Engineering Sciences, Vol. 06, No. 1 (2024) 271-280, doi: 10.24874/PES.SI.24.02.010

RMSE-20%, MAE-10% and R-squared-97% show that might be intricate. Because these composites are so
performs better than conventional approaches. This different, it becomes difficult to predict their mechanical
represents not only a major breakthrough in materials properties. Higher-level insights into the behavior of the
science but also advances the more general objectives of material can be obtained by using sophisticated
sustainability and efficiency in alloy development and characterization methods, such as imaging and in-situ
production processes. With different reinforcing testing. Predictive model improvement and refinement
materials and production methods, alumina composites are possible with this data.

References:

Agrawal, A., & Satapathy, A. (2019). Thermal, mechanical, and dielectric properties of aluminium oxide and solid glass
microsphere‐reinforced epoxy composite for electronic packaging application. Polymer Composites, 40(7), 2573-
2581. doi:https://doi.org/10.1002/pc.25050
Altıparmak, S. C., Yardley, V. A., Shi, Z., & Lin, J. (2022). Extrusion-based additive manufacturing technologies: State
of the art and future perspectives. Journal of Manufacturing Processes, 83, 607-636. doi:
https://doi.org/10.1016/j.jmapro.2022.09.032
Arpitha, G. R., Mohit, H., Madhu, P., & Verma, A. (2023). Effect of sugarcane bagasse and alumina reinforcements on
physical, mechanical, and thermal characteristics of epoxy composites using artificial neural networks and response
surface methodology. Biomass Conversion and Biorefinery, 1-19. doi: https://doi.org/10.1007/s13399-023-03886-7
Banerjee, T., Dey, S., Sekhar, A. P., Datta, S., & Das, D. (2020). Design of alumina reinforced aluminium alloy
composites with improved tribo-mechanical properties:A machine learning approach. Transactions of the Indian
Institute of Metals, 73, 3059-3069. doi: https://doi.org/10.1007/s12666-020-02108-2
Boopalakrishnan, G., Rexliene, M. J., Praveen, R., Balaji, V., Dhandapani, A., & Sridhar, J. (2023). Polyester-Based
Bio-Composites for Marine Applications. In Polyester-Based Biocomposites (pp. 231-252). CRC Press.
Cygan, T., Wozniak, J., Petrus, M., Adamczyk-Cieslak, B., Kostecki, M., &Olszyna, A. (2019). The effect of
microstructure evolution on mechanical properties in novel alumina-montmorillonite composites. International
Journal of Refractory Metals and Hard Materials, 80, 195-203. 10. doi: https://doi.org/10.1016/j.ijrmhm.2019.01.018
Deng, Z. H., Yin, H. Q., Jiang, X., Zhang, C., Zhang, G. F., Xu, B., ... & Qu, X. H. (2020). Machine-learning-assisted
prediction of the mechanical properties of Cu-Al alloy. International Journal of Minerals, Metallurgy and Materials,
27, 362-373. doi: https://doi.org/10.1007/s12613-019-1894-6
Devadiga, U., Poojary, R. K. R., & Fernandes, P. (2019). Artificial neural network technique to predict the properties of
multiwall carbon nanotube-fly ash reinforced aluminium composite. Journal of Materials Research and Technology,
8(5), 3970-3977. doi: https://doi.org/10.1016/j.jmrt.2019.07.005
Devi, M. A., Prakash, C. P. S., Chinnannavar, R. P., Joshi, V. P., Palada, R. S., & Dixit, R. (2020, September). An
informatic approach to predict the mechanical properties of aluminum alloys using machine learning techniques. In
2020 International Conference on Smart Electronics and Communication (ICOSEC) (pp. 536-541). IEEE. doi:
https://doi.org/10.1109/ICOSEC49089.2020.9215277
Goswami, S., Ghosh, R., Hirani, H., & Mandal, N. (2022). Mechano-tribological performance of Graphene/CNT
reinforced alumina nanocomposites–Review and quantitative insights. Ceramics International, 48(9), 11879-
11908. doi: https://doi.org/10.1016/j.ceramint.2022.02.214
Hajilounezhad, T., Bao, R., Palaniappan, K., Bunyak, F., Calyam, P., & Maschmann, M. R. (2021). Predicting carbon
nanotube forest attributes and mechanical properties using simulated images and deep learning.npj Computational
Materials, 7(1), 134. doi: https://doi.org/10.1038/s41524-021-00603-8
Katırcı, R., &Yıldız, B. K. (2023). Prediction of fracture behavior of Al2O3-Cr2O3 ceramics in different Cr2O3 ratios
under flexure load using machine learning methods. Computational Materials Science, 228, 112362. doi:
https://doi.org/10.1016/j.commatsci.2023.112362
Kılıç, H., & Savaş, Ö. (2022). Abrasive Wear Behavior of Centrifugally Cast Al Alloy Composites with TiB2
Reinforcement. Tribology in Industry, 44(3), 461–469. https://doi.org/10.24874/ti.1282.04.22.06
Kordijazi, A., Behera, S., Patel, D., Rohatgi, P., &Nosonovsky, M. (2021). Predictive analysis of wettability of Al–Si-
based multiphase alloys and aluminum matrix composites by machine learning and physical modeling. Langmuir,
37(12), 3766-3777. doi: https://doi.org/10.1021/acs.langmuir.1c00358
Kota, N., Charan, M. S., Laha, T., & Roy, S. (2022). Review on development of metal/ceramic interpenetrating phase
composites and critical analysis of their properties. Ceramics International, 48(2), 1451-1483. doi:
https://doi.org/10.1016/j.ceramint.2021.09.232
Kumar, T. P., Shalini, S., Priyadharshini, G., & Raghu, R. (2022). Parametric optimization of dry sliding wear behavior
of A356 Alloy-Zircon Composites. Tribology in Industry, 44(4), 719–730. https://doi.org/10.24874/ti.1348.08.22.11

279
 Kumar et al., Predicting alumina composites’ mechanical characteristics using a machine learning approach

Lakhdar, Y., Tuck, C., Binner, J., Terry, A., & Goodridge, R. (2021). Additive manufacturing of advanced ceramic
materials. Progress in Materials Science, 116, 100736. doi: https://doi.org/10.1016/j.pmatsci.2020.100736
Liu, J., Zhang, Y., Zhang, Y., Kitipornchai, S., & Yang, J. (2022). Machine learning assisted prediction of mechanical
properties of graphene/aluminium nanocomposite based on molecular dynamics simulation. Materials & Design, 213,
110334. doi: https://doi.org/10.1016/j.matdes.2021.110334
Liu, Y., Wu, Q., Wang, P., & Zhuang, W. (2023). Forming process prediction of a self-piercing riveted joint in carbon
fibre reinforced composites and aluminium alloy based on deep learning. Journal of Manufacturing Processes, 106,
453-464. doi: https://doi.org/10.1016/j.jmapro.2023.10.015
Najjar, I. M. R., Sadoun, A. M., Alsoruji, G. S., Abd Elaziz, M., &Wagih, A. (2022). Predicting the mechanical
properties of Cu–Al2O3 nanocomposites using machine learning and finite element simulation of indentation
experiments. Ceramics International, 48(6), 7748-7758. doi: https://doi.org/10.1016/j.ceramint.2021.11.322
Palmero, P. (2019). Zirconia‐Based Composites for Biomedical Applications.Bioceramics and Biocomposites: From
Research to Clinical Practice, 57-85. https://doi.org/10.1002/9781119372097.ch4
Prusov, E. S., Aborkin, A. V., Deev, V. B., Bokaryov, D., & Babin, D. (2023). Tribological behavior of cast aluminum
matrix composites after multiple remelting. Journal of Materials and Engineering, 1(3), 131–135.
https://doi.org/10.61552/jme.2023.03.006
Rajput, R., Raut, A., &Setti, S. G. (2022). Prediction of mechanical properties of aluminium metal matrix hybrid
composites synthesized using Stir casting process by Machine learning. Materials Today: Proceedings, 59, 1735-
1742. doi: https://doi.org/10.1016/j.matpr.2022.04.316
Raju, S. S., Murali, G. B., & Patnaik, P. K. (2020). Ranking of Al-CSA composite by MCDM approach using AHP–
TOPSIS and MOORA methods. Journal of Reinforced Plastics and Composites, 39(19-20), 721-732. doi:
https://doi.org/10.1177/0731684420924833
Sam, M., Radhika, N., Sidvilash, V., & Mohanraj, T. (2022b). Investigation on the mechanical and wear behaviour of
AL-6082-BN-B4C-CorN COB ASH hybrid composites. Tribology in Industry, 44(1), 294–309.
https://doi.org/10.24874/ti.1165.08.21.11
Sharma, A., Mukhopadhyay, T., Rangappa, S. M., Siengchin, S., & Kushvaha, V. (2022). Advances in computational
intelligence of polymer composite materials: machine learning assisted modeling, analysis and design. Archives of
Computational Methods in Engineering, 29(5), 3341-3385. 10.1007/s11831-021-09700-9
Thirumoorthy, A., Arjunan, T. V., & Kumar, K. S. (2019). Experimental investigation on mechanical properties of
reinforced Al6061 composites and its prediction using KNN-ALO algorithms. International Journal of Rapid
Manufacturing, 8(3), 161-177. doi: https://doi.org/10.1504/IJRAPIDM.2019.100498
VarolÖzkavak, H., Ince, M., &Bıçaklı, E. E. (2023). Prediction of Mechanical Properties of the 2024 Aluminum Alloy
by Using Machine Learning Methods. Arabian Journal for Science and Engineering, 48(3), 2841-2850. doi:
https://doi.org/10.1007/s13369-022-07009-8
Wang, J., Fa, Y., Tian, Y., & Yu, X. (2021). A machine-learning approach to predict creep properties of Cr–Mo steel
with time-temperature parameters. Journal of Materials Research and Technology, 13, 635-650. doi:
https://doi.org/10.1016/j.jmrt.2021.04.079
Wiciak-Pikuła, M., Felusiak-Czyryca, A., & Twardowski, P. (2020). Tool wear prediction based on artificial neural
network during aluminum matrix composite milling. Sensors, 20(20), 5798. doi: https://doi.org/10.3390/s20205798
Yu, Z., Ye, S., Sun, Y., Zhao, H., & Feng, X. Q. (2021). Deep learning method for predicting the mechanical properties
of aluminum alloys with small data sets. Materials Today Communications, 28, 102570. doi:
https://doi.org/10.1016/j.mtcomm.2021.102570

Ashwini Kumar Arunkumar Devalapura Thimmappa Ritesh Kumar

Arka Jain University, Jamshedpur, JAIN (Deemed-to-be University), Maharishi University of
Jharkhand, India Ramanagara District, Karnataka, India Information Technology, Uttar
[email protected] [email protected] Pradesh, India
ORCID 0000-0003-3558-8054 ORCID 0000-0001-8034-1881 [email protected]
ORCID 0009-0008-6351-083X
Manali Gupta
Noida Institute of Engineering &
Technology,
Greater Noida, Uttar Pradesh, India
[email protected]
ORCID 0000-0001-6876-817X

280