STATISTIKA DAN PROBABILITAS

( jurnal Internasional )

“Prediction of Weight Percentage Alumina and Pore Volume Fraction in Bio-Ceramics Using
Gaussian Process Regression and Minimax Probability Machine Regression”

OLEH

AGUNG PRTAMA
EICI19014

JURUSAN TEKNIK MESIN

FAKULTAS TEKNIK
UNIVERSITAS HALU OLEO
KENDARI
2021
1.1 Introduction

In bio-ceramics, the flexural strength is more important for its application in human body
for the reason that the bio-ceramics is used for replacing or repairing bone joints, teeth etc. The
alumina amount present in bio-ceramics plays a vital role in achieving required density, grain size
and flexural limits of required limit values. The pores in bio-ceramics are obtained by burning
foreign combustible products while sintering it. The pores have different sizes and pore volume
fraction is not distributed evenly in it. The pore size variation did not

exhibit much difference in mechanical properties of it, but pore volume fraction had
much influence on the mechanical properties of it.

The porosity has much influence on Poisson’s ratio of the alumina based ceramics,
although Poisson’s ratio is generally considered as constant value, almost in all the Research
works. Many characteristics of alumina based ceramics have been measured by many
experiments. The Fracture Toughness of alumina based ceramics is indurate to volume fraction of
alumina. The experiments have shown that the porous hydroxyapatite ceramics can be fabricated
equivalent to the quality of natural bone by ascertained pores characteriistics such as pore size,
pore volume fraction and pore structures.

The pores in ceramics like Zirconia, have exhibited low thermal conductivity up to the
low thermal conductivity upto the lowest value of 0.1 Wm-1k-1 6. The above porosity and the
pore size distribution can be varied by the process parameters such as quantity of polymer and
sintering temperature7 . The porous characteristics in the ceramics can be fabricated by the size
amount and distribution of pore-forming agents8 . The large pores provided a high porosity
whereas the small pores provided a large surface area9 . Some scientists made some research
works to improve the surface lubrication of the ceramics. When the ceramics doped with CuO
(Copper Oxide) there was considerable reduction of friction10. A mechanism of lubrication of
alumina was determined so that lubricous hydroxide layer was formed and the friction was very
much reduced11. Generally the experimentations in ceramics consume much time and it is a
tedious process. Research scientists Altinkok and Koker have applied Artificial Neural Network
for estimating the output parameters such as weight percentage alumina (wa) and pore volume
fraction (pv) using amount of Silicon Carbide (SiC) as input parameter12. In this article, we use
GPR and MPMR for estimating the performance on the data given by Altinkok and Koker.

1.2 Gaussian Process Regression (GPR)

GPR is a Bayesian process13 and it has been applied in many research works in many
fields for prediction and exhibited good results14-17. GPR is a non-parametric model for
prediction of a function for a given set of data. In GPR, we have to find function ‘f’ with
minimum errors from the data D = {x, y}. A Gaussian process defines a distribution over function
p(f) which can be used for Bayesian Regression

p ( f ) p(d /f )
p ( f /d ) =
p(D)
Gaussian Process Covariance function (Kernels)

In the Gaussian process, it is always assumed that the mean value is always zero. In
Gaussian process y or f(x) is a dependent function of x, for which we are interested to learn in the
form of supervised learning, where the training data values are in control with fine distinctions. In
Gaussian process, one function is considered as main function, which relates the value of
observed one to value of another is called ‘Covariance Function’, generally denoted as k(x, x’).

1.3 Minimax Probability Machine Regression (MPMR)

With respect to GPR, MPMR avoids assumptions based on the data distribution. The
regression model obtained by using MPMR has the error within the assigned error of true
regression model. MPMR uses Mercer’s Kernel for obtaining non-linear regression models.
Many Research scientists have applied MPMR for prediction in various fields18-20.

Let us assume that the unknown regression function f: Rd  R, which has the form

y=f ( x )+ p

where x ∋ Rd and y ∈ R are within the boundary distribution A.

In the above function y, ρ is noise or fitting error and has zero mean value i.e., E(ρ) = 0,
variance Var(ρ) = σ2 and finite σ  R.

Let us consider the training set examples

T =¿

Where  i  {1, 2 …N}, xi = {xi1, xi2 …xid}Rd and yi  R. We have two objectives
one is to find the approximation function ( ) ˆ yˆ  f x and the second objective is to find the error
ε > 0 and fitting error (minimum probability)

  inf Pr{| yˆ  y | }

The MPMR formulation for the above approximation is given by

N
^y = f^ ( x )=∑ β i K ( x i , x ) +b
i=1

where k(xi, x) = φ(xi).φ(x) is a Kernel which satisfies Mercer’s conditions xi  i {1,2,

….N} are obtained from the learning data Γ . In the above formulation βi and b  R are the
outputs of MPMR learning algorithm.

1.4 Results and Discussions

Generally in machine learning models, 70% of the data sets are considered training set
and 30% meant for test set. Using 70% of the original data, a prediction function is estimated and
the same function is tested for validity with the remaining 30% of the data set. Here, the co-
efficient of correlation (R) is the measure of the robustness of the machine learning model. The R
value should be closer to ‘1’; the predicted function of the machine learning theory model is more
robust in nature.

In this work, 18 data sets are taken from the source data, in which 12 data sets are
considered for training data sets and the remaining 6 data sets are considered for testing sets.

In Fig.1, the GPR model is applied for prediction of weight percentage alumina. In
training data sets, the R value is found to be 0.999 and in testing data sets R value is found to be
0.998. In the above diagram, almost all the training data sets lie on the line which shows actual
value is equal to predicted value, but the 2 data values test set lie slightly away from the line. This
shows that R value of testing data set is slightly lower than R value of the training data set.

Fig. 1 - Performance of GPR model on weight percentage alumina (wa).

In the Fig.2, the GPR model is applied for estimation of pv. The value of R of training
data set (R=0.998) is equal to the value of R of testing data, but comparing with previous fig. 1,
the R value of GPR on training data of pv (R=0.998) is slightly lesser than R value of GPR
(R=0.999) on wa. In the fig.2 the training data set and testing data set is not much closer to the
actual-prediction line as in Fig.1.
 Fig. 2 - Performance of GPR model on pore volume fraction (pv).

In the Fig.3, the MPMR model is applied for estimation of wa. Comparing with the fig.1,
the R values of MPMR (R = 0.998 for training data and R = 0.996 for testing data) are lower than
the R values of GPR (R = 0.999 for training and R = 0.998 for testing data sets). The R value of
MPMR is marginally lesser than R value of GPR on wa. This shows that GPR performs much
better than MPMR in estimation of wa. Even though MPMR is comparatively as good as GPR,
but GPR outperforms MPMR in estimation of wa.

Fig. 3 - Performance of MPMR model on weight percentage alumina (wa).

In the Fig.4, the MPMR model is applied for estimation of pv. Comparing with fig.2, the
MPMR performance over GPR is comparatively low, i.e., marginally. Here, R values of MPMR
on pv (R = 0.998 for training data and R = 0.996 for testing data) are lesser than R values of GPR
values on pv (R = 0.998 for both training and testing data). Even though the robustness of GPR
and MPMR on training data remains the same, the robustness on testing data sets differs

Fig. 4 - Performance of MPMR model on pore volume fraction (pv).