**Gaussian Process Methodology for Multi-Frequency Marine Controlled-Source Electromagnetic Profile Estimation in Isotropic Medium**

#### **Muhammad Naeim Mohd Aris 1,\*, Hanita Daud 1, Sarat Chandra Dass <sup>2</sup> and Khairul Arifin Mohd Noh <sup>3</sup>**


Received: 18 August 2019; Accepted: 19 September 2019; Published: 27 September 2019

**Abstract:** The marine controlled-source electromagnetic (CSEM) technique is an application of electromagnetic (EM) waves to image the electrical resistivity of the subsurface underneath the seabed. The modeling of marine CSEM is a crucial and time-consuming task due to the complexity of its mathematical equations. Hence, high computational cost is incurred to solve the linear systems, especially for high-dimensional models. Addressing these problems, we propose Gaussian process (GP) calibrated with computer experiment outputs to estimate multi-frequency marine CSEM profiles at various hydrocarbon depths. This methodology utilizes prior information to provide beneficial EM profiles with uncertainty quantification in terms of variance (95% confidence interval). In this paper, prior marine CSEM information was generated through Computer Simulation Technology (CST) software at various observed hydrocarbon depths (250–2750 m with an increment of 250 m each) and different transmission frequencies (0.125, 0.25, and 0.5 Hz). A two-dimensional (2D) forward GP model was developed for every frequency by utilizing the marine CSEM information. From the results, the uncertainty measurement showed that the estimates were close to the mean. For model validation, the calculated root mean square error (RMSE) and coefficient of variation (CV) proved in good agreement between the computer output and the estimated EM profile at unobserved hydrocarbon depths.

**Keywords:** multiple frequency marine controlled-source electromagnetic technique; Gaussian process; uncertainty quantification; computer experiment, electromagnetic profile estimation

#### **1. Introduction**

Nowadays, the controlled-source electromagnetic (CSEM) technique is a significant application to detect and discover hydrocarbon-filled reservoirs based on the principles of electromagnetic (EM) propagation. For decades, CSEM application has been widely exercised in onshore geophysical exploration (e.g., [1,2]). The efficiency of this application to characterize offshore hydrocarbon reservoirs has also been proven by many oil and gas companies around the world. Li and Key [3] stated that at early stage, marine CSEM application was employed to study the electrical conductivity of the upper mantle and oceanic crust (e.g., [4–8]). Studies related to the commercial application of the marine CSEM technique in offshore hydrocarbon exploration can be found in [9–17]. Previously, the seismic sounding survey, which employs acoustic waves, was solely utilized to map geological structures that have different acoustic properties [18]. This survey was very important to hydrocarbon exploration due to its capability of providing information of the subsurface. According to [19], seismic data interpretation provides good resolution of the subsurface structures underneath the seabed; however, it has deficiencies. It is said that seismic surveys are unable to distinguish the fluid content inside the reservoirs, whether brine (conductive seawater) or hydrocarbon. Zaid et al. [18] mentioned that seismic sounding is not compatible to the direct detection of the pore fluid reservoirs. Note that EM and seismic techniques are sensitive to two different properties of subsurface; thus, the marine CSEM technique was developed as a complementary interpretational tool to specifically characterize the target reservoirs.

The marine CSEM technique also is referred to as a seabed logging (SBL) application. This is thoroughly described by [10]. This application is particularly able to reduce ambiguities in data interpretation in hydrocarbon exploration. Andreis and MacGregor [16] stated that by studying the reflected EM signal, resistive mediums such as hydrocarbon, gas, and hydrate can be discovered to depths of several kilometers from the seabed. In addition, the resistivity of the subsurface in offshore environments is commonly identified by robust anisotropy because of the sedimentation factor [20]. Note that for a medium that is horizontally stratified, the subsurface is generally less resistive in the horizontal (parallel) direction than in the vertical (perpendicular) direction [20]. Offshore hydrocarbon reservoirs are normally embedded in a high conductive medium unlike the common case of onshore hydrocarbon reservoirs [21]. Hydrocarbon-filled reservoirs are known to have very high electrical resistivity compared to its surroundings, such as of saline water and sedimentary rocks. These structures are very conductive. From [22], hydrocarbon is known to have electrical resistivity between 30 and 500 Ohm meter, whereas the resistivity values of seawater and sediment are 0.5–2 Ohm meter and 1–2 Ohm meter, respectively. If a target reservoir is brine saturated, it is normally a few orders of magnitude less electrically resistive than a hydrocarbon-filled reservoir. From these characteristics, the resistivity of the subsurface can be resolved via data of electric (E-) and magnetic (H-) fields obtained from the marine CSEM survey. The measurement of amplitude and phase of Eand H- fields can be utilized to determine the geological subsurface. Li and Key [3] mentioned that the amplitude and phase of an EM field will vary depending on the resistivity of the structures beneath the seabed, the depth of seawater, and the source–receiver offset.

In the marine CSEM technique, data collected can be interpreted in two different groups depending on the domain—either in time-domain or frequency-domain. Analyzing data in timeor frequency-domains would theoretically give the same output/information [23]. Reyes et al. [24] mentioned that frequency-domain marine CSEM is normally used for the case of oil prospecting. In the frequency-domain application, an antenna/transmitter of towed EM dipole is generally used to generate a low-frequency EM field, and returned/reflected signals recorded by receivers placed on the seabed are utilized for resistivity distribution analysis. The choice of frequency is very crucial in marine CSEM application. In a standard configuration, marine CSEM surveys use a deep-towed horizontal electric dipole (HED) transmitter to emit a low-frequency EM wave which is usually between 0.1 Hz and 10 Hz to an array of seabed receivers, and normally, the transmitter is towed at 30–50 m above the seabed [25]. Practically, the low-frequency of EM waves is used in a deep water environment as the signal transmission due to the fact that low-frequency is able to yield farther penetration through seawater columns into sedimentary rocks. Next, for the receiver, there are two types of receiver configurations—inline and broadside. Inline configuration is when the separation distance is parallel to the direction of antenna, whereas broadside is when the source–receiver offset is normal to the antenna's direction [26]. Due to the characteristics of subsurface conductivity, EM signals spread with a higher rate through the seafloor than through the seawater. The EM energy, which is transmitted from the towed-source, spreads in all directions and is quickly attenuated in conductive medium such as sediment. The occurrence of all possible signal contributions is depicted in Figure 1.

**Figure 1.** Basic layout of marine CSEM application in hydrocarbon exploration. The source is towed nearly to the seabed receivers. A target reservoir is expected to be embedded in the conductive sedimentary rocks. The EM signal spreads in all direction through the seawater, air–seawater interface, and sediment before being recorded by the EM receivers.

Based on Figure 1, the straight-line arrow denotes the direct-wave travelling from the source to the seabed receiver without any interaction with the geological subsurface beneath the seabed. Researchers in [27,28] mentioned that direct-wave dominates the data collection at short source–receiver separation offsets. Next, the dotted-line arrow represents the reflected and refracted wave from the source upwards to the air–seawater interface and vertically going back through the seawater to the receiver (i.e., air-wave). This wave travels with high rate of velocity (propagates with no attenuation) to the water surface since air is an infinitely electrical resistive medium. Seawater depth can influence the measured EM signal. Air-wave contribution increases as the depth of seawater decreases. Weiss [29] asserted that this contribution becomes significant in a seawater depth of roughly less than 300 m. Both these signals, direct- and air- waves, do not contain any information about hydrocarbon-filled reservoirs. Last signal contribution is denoted as dashed-line arrow. It is known as the reflected and refracted wave (i.e., guided-wave). This wave diffuses outward from the source through seawater column and then through the high resistive formations with less attenuation. The transmitted wave has to enter the formations at certain angle which is between 0◦ and 11◦ in order to set the guided mode [28]. This reflected and refracted wave strongly dominates the recordings at intermediate source–receiver offsets (~3 to ~8 km). The detection of the guided-wave is the basis of the marine CSEM survey.

In the context of geophysics forward modeling, electrical resistivity has an important role in oil and gas exploration. Numerical modeling is a crucial component that provides information of the electrical resistivity of the subsurface. There are various computational techniques exercised in EM applications such as the Finite Element (FE), the Finite Difference (FD) and the Method of Moment (MOM) [30], while researchers in [20] have said that FE, FD and integral equation (IE) methods are among the most famous numerical techniques for modeling EM data. According to [20], the FE method is more reliable for EM forward modeling in a complex geological structure when compared to FD and IE methods. Based on the literature, FE is a usual numerical technique exercised in CSEM modeling for hydrocarbon exploration. This computational method uses unstructured grids that can be easily conformed to irregular boundaries compared to the FD method. The MOM is less preferred as well in

marine CSEM data interpretation since this method produces more complex derivations of governing equations than the FE method. The traditional FD method is easier to implement and maintain than the FE method, but the method is based on structured grids. This means that grid refinement is not possible, and hence it affects the overall computational processes [31]. Unstructured grids have long been exercised in various fields such as engineering and applied mathematics; however, this feature has only recently been used in the EM geophysical field, as exemplified by the use of FE method code in marine EM surveys (e.g., [3,32–34]). This feature can realistically replicate the complexities of geological structures [19].

Even though this numerical technique is very powerful, the ad hoc design of meshes in FE is time-consuming. Li and Key [3] mentioned that the most time-consuming task in their code (FE algorithms) is the solutions of the linear equation systems. The study could take a very lengthy computational time if they used all wavenumbers in the mesh refinement for a full solution. Bakr et al. [35] also stated that the most time-consuming tasks in FE are evaluating the integrals and solving the linear equations. For typical simulations, a few million elements are involved in the linear equation systems [31]. According to [24], the execution of real-field simulations in EM problems needs the use of high-performance computing (HPC). This is because typical actual executions require more than hundred thousand realizations which involve millions of degrees of freedom for each process. The computational and memory requirements to solve such solutions may become a serious challenge. It can be more complicated for higher-dimensional EM forward modeling and inversion. Besides forward modeling, inversion is also a powerful way to recover the electrical conductivity profile beneath the seabed given measurements of EM fields acquired from real-field surveys. Not to mention, nowadays, inverse modeling comes with robust inversion schemes and incorporation of more procedures and measurements. This makes it possible to compute the EM fields at the seabed receivers precisely and provide accurate geometry resolution. However, it is said that inversion algorithms tend to be computationally expensive due to the forward modeling schemes. Indeed, an inversion process needs multiple EM forward solutions [35]. Furthermore, in terms of application, the marine CSEM technique generates huge amounts of data (captured by seabed EM receivers with a moving HED source); therefore, processing those data has become a challenging task to many geophysicists [9,36]. Modeling the marine CSEM data is related to the need of accurate representation of very complex geo-electrical models, and the algorithms used should be powerful and fast enough to be applied to repeated use of hundreds of iterations and multiple source–receiver positions. In addition, understanding the noisy CSEM data to quantify the uncertainties involved in EM modeling also is very crucial. Constable and Srnka [14] stated that the economic challenges (e.g., related to drilling) increase as the hydrocarbon exploration moves to deeper offshore environments. Thus, any additional data that can be obtained or collected will be advantageous to the exploration if there is a potential to de-risk a given expectation [19].

In order to seek the most favorable balance between the computational cost involved in the interpretation of EM geophysical data and the accuracy of the modeling, our interest is focused on processing the one-dimensional (1D) frequency-domain marine CSEM data using Gaussian process (GP) algorithms. We propose GP as a methodology for two-dimensional (2D) forward modeling of the marine CSEM technique to provide information on EM profiles when hydrocarbon is present at various depths in isotropic mediums. This forward modeling provides the uncertainty measurement of the estimation in terms of variance. Although the existing CSEM models (the existing numerical modeling techniques) provide robust representation of real-field models, this work has significant contribution for hydrocarbon detection as well. This attempt is very useful and helpful when collected sets of data in CSEM surveys are insufficient for the interpretation of higher-dimensional modeling and inversion. This analysis also could reduce time in the CSEM workflow since forward GP modeling is able to provide uncertainty quantifications without integrating or combining any other numerical quantifiers. On top of that, there is its simplicity, as GP only involves simple equations which means faster computation which only needs basic memory space. Note that this analysis utilizes the simulation datasets generated through a commercial software, namely, Computer Simulation Technology (CST). Information of the CST software can be found in [37]. The details and literature of GP application are thoroughly elaborated in next section.

#### **2. Statistical Background: Gaussian Process in Computer Experiments**

Gaussian Process (GP) is random function which has a property that any finite number of evaluations of the (random) function has a multivariate Gaussian distribution. GP is fully specified by a mean function, *m*(*x*), and a covariance function, *k*(*x*, *x* ). The Gaussian distribution has mean and covariance values in the forms of vector and matrix evaluations, respectively [38]. Here, *x* represents all potential independent variables that influences the outputs/responses. GP is a non-parametric and probabilistic method for fitting functional forms based on domain observations. It differs from most of other black-box identification approaches where it does not approximate the modelled system by fitting the parameters of basis function, but rather searches for relationships among the measured data. This non-parametric regression method does not need a fixed discretization. This technique is able to provide predictive mean values and uncertainty of the estimation measured in terms of variance. This variance reflects the quality of the output/information. It is an important numerical measure when it comes to distinguishing GP from the other computational intelligence methods. According to [38], GP is suitable for modeling uncertain processes or data which are unreliable, noisy or contain missing values. GP has been used in many different applications. Studies related to application of GP in various fields can be found in [39–45]. In general, prior belief of spatial smoothness is specified through a covariance defined by similarity characteristics. Training observations are then considered as the realizations from the updated multivariate Gaussian (i.e., posterior). Thus, the conditional realizations from the posterior are simply the testing output at all untried or unobserved points. The mathematics behind this concept are thoroughly explained in the methodology section.

Computer experiments are well-known and not new in science, technology, and engineering. This medium is getting very popular for solving scientific and engineering problems. Nowadays, scientists prefer to use computer simulators rather than doing case studies or conducting any related physical experiments. It can be implemented in any circumstances including experiments that are impossible to do physically, with shorter time taken than in the real situation. Computer experiments are run by means of a complex code and highly developed theories of physics, mathematics, and engineering fields. Sacks et al. [46] described that experimenters usually aim to estimate/predict the output at unobserved input points, optimize the function of the output points, and calibrate the computer code to physical data. To this end, [46] and the subsequent works modelled the output of a computer model, *Y*(*x*), based on input *x* as a sum of regression terms, β*j*, and stochastic component, *Z*(*x*). *Y*(*x*) is defined in Equation (1).

$$\mathcal{Y}(\mathbf{x}) = \sum\_{j=1}^{k} \beta\_j f\_j(\mathbf{x}) + Z(\mathbf{x}),\tag{1}$$

where *fj*(*x*) is a known function with *j* = 1, 2, ... , *k*, and *Z*(*x*) is a random process with a zero-mean and a covariance. The most famous choice of the stochastic component, *Z*(*x*), is a GP where the distribution of the GP is assumed to be a normal (Gaussian) distribution with a mean and a covariance function. According to [47], GP is used as the surrogate model for any complex mathematical models which consume a lot of time to solve. GP is flexible in representing the computer output, *Y*(*x*), and it is feasible to obtain analytical formulas of the predictive distribution and to design the equations.

From the reported literature, there are a few studies calibrating CST computer output of marine CSEM applications with GP (e.g., [48,49]). However, these studies present 1D GP modeling of SBL applications where they only considered univariate independent variables. Besides, the work only focused on predicting the presence of the hydrocarbon layer at a known depth, while we consider various depths of hydrocarbon at observed and unobserved depth levels. This is because the location of hydrocarbon reservoirs can be anywhere and is uncertain in real-field environments. Aris et al. [50] also described forward GP modeling of SBL applications, but the paper only focused on

one transmission frequency and no error was considered in the presented GP modeling. Since CST computer output is assumed to generate very clean data, considering the error in modeling is very important to marine CSEM data processing. Thus, this attempt is novel in two ways; first, it proposes GP methodology to process marine CSEM data calibrated with CST computer output at multiple transmission frequencies at which hydrocarbon is present at all possible depth levels (250–2750 m); second, this is a data-dependent analysis where it utilizes an uncertainty quantification provided by the GP in marine CSEM data processing with error considerations before in-depth analysis. This may enhance the EM data interpretation where the EM profile is estimated with the measurement of variance at various possible depths of the hydrocarbon layer, which helps decision-making for hydrocarbon detection in marine CSEM applications.

#### **3. Methodology**

The methodological flow is based on a three-step procedure; (i) synthetic seabed logging (SBL) modeling using Computer Simulation Technology (CST) software, (ii) developing two-dimensional (2D) forward Gaussian Process (GP) models for multiple EM transmission frequencies, and (iii) model validation using the root mean square error (RMSE) and the coefficient of variation (CV).

#### *3.1. Synthetic SBL Modeling Using CST Software*

We designed synthetic models of typical marine CSEM application for hydrocarbon exploration which have various depths of hydrocarbon at multiple frequencies by using CST software. The transmissions were tested at frequencies of excitation current of 0.125, 0.25, and 0.5 Hz. Note that in the CST software, Maxwell's equations are discretized using the Finite Integration Method (FIM). FIM solves the Maxwell's equations in a finite calculation domain in grid cells to probe the resistivity contrast. For this study, the SBL model is a three-dimensional (3D) canonical structure, which consists of background layers such as air, seawater, and sediment. The model is designed with an air–seawater interface at z = 300 and the seawater thickness is fixed at 1000 m (deep offshore environment). A 200 m-thick horizontal resistive layer (hydrocarbon) is embedded in the sediment layer with various depths from the seabed. The thickness of sediment above the hydrocarbon layer, known as overburden thickness, is varied from 250 m to 2750 m with an increment of 250 m each. The thickness of the overburden layer indicates the depth of hydrocarbon reservoir. Thicker overburden layers mean a deeper location of the hydrocarbon. Both the background and hydrocarbon layers are considered as isotropic. Figure 2 shows the stratified illustration of the horizontal layers used in this study.

**Figure 2.** Illustration of SBL models used in this study. The thicknesses of air, seawater, and hydrocarbon are 300, 1000, and 200 m, respectively. The depth of the hydrocarbon layer (thickness of overburden layer) is varied from 250 m to 2750 m with an increment of 250 m. The total height and length of the models are 5000 and 10,000 m, respectively.

The replication of the 3D structures of the SBL model (length, x: 10,000 m; width; y: 10,000 m; height, z: 5000 m) can be referred to in [50]. The electrical conductivities of air, seawater, sediment, and hydrocarbon are tabulated in Table 1. Their properties are taken from [48].

**Table 1.** Electrical conductivity of every layer considered in the SBL models. Conductivity is the reciprocal of resistivity. The electrical conductivity of the hydrocarbon layer is lower than its surroundings, which are seawater and sediment. It means that the hydrocarbon layer is parameterized with reliable electrical resistivity.


The EM signal is transmitted by a HED source located with the orientation of *x*-direction in the seawater with coordinates (5000, 5000, 1270). This means that the inline transmitter pointing along x-axis is positioned at *x*, *y* = 5000, and a height of 30 m above the seabed. The HED source is held stationary at the center of the model. The values of the EM field are measured along an inline profile through the SBL model. In this study, the current strength of the HED source is fixed at 1250 A. The source–receiver separation distances (offsets) are varied along the replication model. An array of 1000 seabed receivers is placed along the seabed at *x* ranges of 0–10,000 m and y = 5000 m. This means, receivers are positioned along the seabed for every 10 m from 0–10,000 m of x-orientation. As a demonstration, Figure 3 is the mesh view of the replication of the marine CSEM model in isotropic medium, replicated by the CST software at a hydrocarbon depth of 250 m.

**Figure 3.** Mesh view of the 3D SBL model at hydrocarbon depth of 250 m replicated by CST software. Both background and hydrocarbon layers are set as isotropic. The hydrocarbon layer is designed with length (x) of 10,000 m, width (y) of 5000 m and height (z) of 200 m.

#### *3.2. Developing 2D Forward GP Models at Multiple EM Transmission Frequencies*

Let *Y*(*xi*) be the CST computer output at *k* different input specifications for every frequency used (0.125, 0.25, and 0.5 Hz), where *i* = 1, 2, ... , *k*. The input variable, *x*, can be univariate

or multivariate, but in this study, we exercise a bivariate independent variable. Source–receiver separation distance (offset), *s*, and depth of hydrocarbon layer from the seabed, *h*, are considered as input variables where *x* = (*s*, *h*). In this paper, we focus on processing non-normalized CST output, *Y*(*x*), which is the magnitude of the E-field (amplitude) obtained from static source–receiver combinations where the transmitter is fixed at the center of the SBL model. For every frequency, we have source–receiver separation distances, *si* = {*i* = 1, 2, ... , 210}, and hydrocarbon depths, *hi* = {*i* = 1, 2, ... , 11}, which are from 250 to 2750 m with and increment of 250 m each. Thus, in this study, we have *k* = 210 × 11 × 3 = 6930 different input specifications of CST computer output that are to be processed.

As mentioned earlier, GP is completely defined by a mean function, *m*(*x*), and a covariance function, *k*(*x*, *x* ). The GP model on function *f* with a zero-mean function, *m*(*x*) = 0, and a covariance function, *k*(*x*, *x* ), can be written as

$$f(\mathbf{x}) \sim G(0, k(\mathbf{x}, \mathbf{x}')). \tag{2}$$

An appropriate correlation function for our GP is selected. The choice of the correlation function is very crucial in computer experiments. It governs the smoothness of the sample path realizations of the GP and is dictated by CST computer output. We choose a popular covariance function which is the squared exponential (SE) function. This covariance function has been widely used in many applications of GP regression and it produces smooth functional estimates. The SE is defined in Equation (3).

$$k(\mathbf{x}, \mathbf{x}') = \sigma\_f^2 e^{\left(\frac{-\left|\mathbf{x} - \mathbf{x}'\right|^2}{2\ell^2}\right)},\tag{3}$$

where σ*<sup>f</sup>* and are signal variance and characteristic-lengths scale, respectively. These hyper-parameters need to be properly estimated, and this is usually by optimizing the marginal likelihood. By referring to Bayes' theorem, we assume that very little prior knowledge about these hyper-parameters are known, and this prior knowledge corresponds to the maximization of marginal log-likelihood. We have three different datasets (three different frequencies). For every dataset, there are 210 data points consisting of offset and hydrocarbon depths as the independent variables corresponding to the magnitude of the E-field as the dependent variable. According to [39], two-thirds of the total data should be considered as training data points and the remainder will be the testing points. Thus, in this paper, for every three data points, the first two data are set as the training data points. This procedure was implemented for every frequency used.

The GP regression model is assumed to generally have a relationship of the form *yi* = *f*(*xi*) + ε where the prior joint distribution for the collection of random variables consist of training and testing points are defined as below

$$\mathbb{E}\left[\begin{array}{c} m\\ m\_{\ast} \end{array}\right] \sim \mathcal{G}\left(\mathbf{0}, \begin{bmatrix} \ K\_{\varepsilon} & K\_{\ast} \\ K\_{\ast}^{T} & K\_{\ast\ast} \end{bmatrix}\right). \tag{4}$$

The vector *m* ∈ *ntrain* is observed at spatial locations *x* ∈ *ntrain*×*nd* where *x* = (*s*, *h*). *ntrain* denotes the number of training data points generated from CST software, while *nd* is the spatial dimension exercised in this paper. Note that for every hydrocarbon depth, *ntrain* = 140 and *nd* = 2. Next, *m*<sup>∗</sup> ∈ *ntest* is a vector that specifies the predicted values at particular spatial locations *<sup>x</sup>*<sup>∗</sup> <sup>∈</sup>*ntest*×*nd* where *x*<sup>∗</sup> = (*s*∗, *h*∗). *ntest* is the number of all desired observations (i.e., testing data points) where *ntest* = 70 for each depth. With 140 data per depth, a matrix *K* ∈ *ntrain*×*ntrain* is defined using Equation (3) for all pairs involved in the training points. Then, *K*<sup>ε</sup> ∈ *ntrain*×*ntrain* is determined such that

$$K\_{\ell} = K + \sigma^2,\tag{5}$$

where <sup>σ</sup><sup>2</sup> represents a diagonal covariance matrix of the specified additive noise at 5%. *<sup>K</sup>*<sup>∗</sup> <sup>∈</sup>*ntrain*×*ntest* , and *K*∗∗ ∈ *ntest*×*ntest* are calculated using Equation (3) where these matrices define the correlation of

training–testing data points and testing–testing data points, respectively. Hence, *m*<sup>∗</sup> is predicted at 70 locations of *x*<sup>∗</sup> per depth. Here, matrix *K*∗∗ only has input from testing data points and it is derived from prior information. Thus, the posterior conditional GP as in Equation (1), given the information of *x*∗, *x* and *m*, is written as

$$p(m\_\*|\mathbf{x}\_\*, \mathbf{x}\_\*|m) = G(m\_\*|\mu\_\*, \Sigma\_\*) \tag{6}$$

Based on the theorem, for every frequency, the Gaussian probability for the random variable *m*<sup>∗</sup> with mean, μ<sup>∗</sup> (i.e., the estimated EM profile at observed and unobserved depths of hydrocarbon), and variance, Σ<sup>∗</sup> (uncertainty measurement in terms of ± two standard deviations), are defined in Equations (7) and (8), respectively. These equations are the main equations of GP regression.

$$
\mu\_\* = \, \prescript{}{\mathbf{K}}{\mathbf{K}}\_{\ast} \prescript{}{\mathbf{K}}{\mathbf{m}}\_{\varepsilon} \, \prescript{}{\mathbf{m}}{\mathbf{m}} \tag{7}
$$

$$
\Sigma\_{\star} = K\_{\star \star} - K\_{\star}^T K\_{\iota}^{-1} K\_{\star} \tag{8}
$$

#### *3.3. Model Validation Using RMSE and CV*

To validate our forward GP model, we calculated the root mean square error (RMSE) and coefficient of variation (CV) of the difference between data predicted by GP (estimates) and the data acquired from CST software. RMSE is able to calculate the difference between an estimate and the true value (observation) corresponding to the expected value of root squared loss. CV is calculated as well to evaluate the relative closeness between true values and estimates in percentage. Two random unobserved depths of hydrocarbon which are 900 m and 2200 m were selected for demonstration purposes. The SBL models with these depths of hydrocarbon were simulated separately. The CST computer output was considered as the true values, and the estimate values are the data predicted by GP at the same depths of hydrocarbon. The RMSE and CV between the true values, y*<sup>i</sup>* , and the estimate values, *y*∗ *i* , are defined as below

$$RMSE = \sqrt{\frac{\Sigma \left(\mathbf{y}\_i - y\_i^\*\right)^2}{a}},\tag{9}$$

$$CV = \frac{RMSE}{\left| \mu\_{y\_i^\*} \right|} \times 100\% \tag{10}$$

where *a* is the total number of testing data points, and μ*y*<sup>∗</sup> *<sup>i</sup>* is the absolute average of *y*<sup>∗</sup> *i* . The estimates from the 2D forward GP model should match the simulation data acquired from the CST software at all observed and unobserved depths very well up until larger offset distances.

#### **4. Results and Discussion**

We implemented this analysis by using GP algorithms in a MATLAB code (built-in function) referred from [51]. We considered a typical synthetic frequency-domain marine CSEM study with three different transmission frequencies (0.125, 0.25, and 0.5 Hz), a HED source, and an array of receivers. Note that for every depth of hydrocarbon, the simulation was simultaneously run for the three transmission frequencies (three datasets per simulation). Every simulation process took ~15 min to generate the three different datasets. Hence, total computational time for the CST software to compute the EM fields for 11 hydrocarbon depths was approximately 165 min (~2 h and 45 min). In addition, in order to make the data interpretable, a logarithmic scale with base 10 was applied to the magnitude of the E-field for every frequency since it involves very small values. Figures 4–6 are the CST computer output at all input specifications for every frequency.

**Figure 4.** Log10 of magnitude of electric field at 0.125 Hz versus source–receiver separation distance (offset). Different hydrocarbon depths yield different EM responses. The offset is from 0 m (left of the SBL model) to 10,000 m (right of the SBL model).

**Figure 5.** Log10 of magnitude of electric field at 0.25 Hz versus source–receiver separation distance (offset). Different hydrocarbon depths yield different EM responses. The offset is from 0 m (left of the SBL model) to 10,000 m (right of the SBL model).

**Figure 6.** Log10 of magnitude of electric field at 0.5 Hz versus source–receiver separation distance (offset). Different hydrocarbon depths yield different EM responses. The offset is from 0 m (left of the SBL model) to 10,000 m (right of the SBL model).

From the results, the replicated SBL model is able to reflect the typical synthetic simulation model of EM application. The simulated datasets resulting from the CST software were in a good agreement with the behavior of real-field CSEM data on the effect of the source–receiver separation distance and variations of hydrocarbon depth to the magnitude of the E-field (amplitude). The strength of E-field is inversely proportional to the source–receiver offset and depth of hydrocarbon. If the source–receiver separation is placed further apart and the hydrocarbon layer is located deeper beneath the seabed, the E-field strength significantly decreases. The acquired responses vary in frequency as well. Here, based on figures above, the EM responses are symmetrical. The EM wave was transmitted from the source which was located at the center of the SBL model. The signal travelled equidistant from the source to the boundaries of the model (left and right of the SBL model). Due to this symmetrical setting, only data from 5000 to 10,000 m were considered for processing purposes. Next, from the figures as well, we can see that the magnitudes of the E-field for all hydrocarbon depths are indistinguishable (especially in Figure 6) at source–receiver offset smaller than ~7400 m. This happens because high transmission frequencies have high attenuation, thus the signal is not able to propagate farther than low-frequency EM wave. Thus, we generalized this analysis by utilizing data for the offset from ~7400 to ~9500 m.

From the CST computer output, we developed a 2D forward GP model for every frequency to provide EM profiles at the observed and unobserved depths of the hydrocarbon. Even though the offset distances considered in this paper are from ~7400 m to ~9500 m, the GP models were set to distances from ~2400 to ~4500 m in order to make it easy to interpret, since the EM signal was transmitted from *x* = 5000 (center of the SBL model). The 2D forward GP models for frequencies of 0.125, 0.25, and 0.5 Hz are depicted in Figures 7–9, respectively.

**Figure 7.** Contour plot of EM profiles (amplitude) for various offset distances (~2400 to ~4500 m) and hydrocarbon depths (250–2750 m) at a frequency of 0.125 Hz with 15 labels.

**Figure 8.** Contour plot of EM profiles (amplitude) for various offset distances (~2400 to ~4500 m) and hydrocarbon depths (250–2750 m) at a frequency of 0.25 Hz with 15 labels.

**Figure 9.** Contour plot of EM profiles (amplitude) for various offset distances (~2400 to ~4500 m) and hydrocarbon depths (250–2750 m) at a frequency of 0.5 Hz with 15 labels.

From the figures, we can see that the GP models are able to provide the information of EM profiles which are the magnitude of the E-field at all desired depths of hydrocarbon (observed and unobserved). Since variance is quantified in GP estimation, we tabulate the average of the variance of EM profiles for all observed depths of hydrocarbon in Table 2 to determine how far the data points are spread out from the mean value.


**Table 2.** The average of variance of EM responses (frequencies: 0.125, 0.25, and 0.5 Hz) at all tried depths of hydrocarbon (250–2750 m with an increment of 250 m each).

The confidence interval exercised by this paper is 95% of the data which lies within ± two standard deviations of the mean. Small values of variance indicate that the data points tend to be very close to the mean. Based on Table 2, all values of the average variances are very small. This implies that the 2D forward GP model is capable of fitting the marine CSEM data very well. Next, for better visualization, we depict a combination of 3D surface plots of the developed 2D forward GP models for every frequency in Figure 10.

**Figure 10.** Combination of 3D surface plots of the 2D GP models for frequencies of 0.125, 0.25, and 0.5 Hz. The x-axis is the offset which is the source–receiver separation distance, the y-axis denotes the depth of hydrocarbon (250–2750 m), and the z-axis represents the log10 of magnitude of the electric field.

We determined the reliability of these 2D forward GP models in providing the information of EM profiles by calculating the RMSE and the CV between true (data generated through the CST software) and estimate values (data from the forward model) at unobserved/untried depths of hydrocarbon. In this section, random unobserved hydrocarbon depths (900 m and 2200 m) were selected. The EM profiles from these depths were compared with the CST computer output at the same depth levels. The RMSE and CV of the EM profiles at both depths are tabulated in Table 3.


**Table 3.** RMSE and CV analyses between EM profiles modelled by GP and EM profiles generated through CST software for all frequencies.

Based on Table 3, the RMSE values obtained are very small and all the CVs are generally less than 1%. This means that the modeling results of the 2D forward GP models are in good agreement with the responses acquired from the CST software even at the unobserved/untried depths of hydrocarbon.

#### **5. Conclusions**

We proposed a methodology of processing marine CSEM data using a statistical approach, Gaussian Process (GP). Based on the results, the EM responses estimated by GP are well fitted with the data generated from the CST software. The results (variance) proved that our proposed 2D forward GP model calibrated with computer simulation output is reliable for marine CSEM data-processing. In general, this 2D forward GP model, which contains EM profiles at various hydrocarbon depths, can be compared to surveyed data, and whichever estimate best matches the data measured from the survey will be the more likely case. The importance of this work lies in the application of GP methodology in multiple frequencies marine CSEM technique by developing a data-dependent model with uncertainty quantification to analyze the EM profiles and understand the geological structure

underneath the seabed. It is too risky to directly make a decision in hydrocarbon exploration without any additional analysis. There are too many challenges involved especially when it comes to deeper offshore environments. Therefore, this methodology should be a data-processing tool that provides beneficial information to hydrocarbon exploration using marine CSEM techniques by utilizing the prior information obtained from real-field data before further analysis.

**Author Contributions:** Conceptualization, H.D. and S.C.D.; methodology, M.N.M.A.; software, H.D.; validation, S.C.D. and K.A.M.N.; writing—original draft preparation, M.N.M.A.; writing—review and editing, M.N.M.A. and K.A.M.N.; supervision, H.D., S.C.D. and K.A.M.N.; funding acquisition, H.D.

**Funding:** This research work was funded by International Grant (cost center: 015-ME0-012).

**Acknowledgments:** We would like to thank Universiti Teknologi PETRONAS for the Graduate Research Assistantship (GRA) Scheme. We are really grateful to have an open access of GPs algorithms, Gaussian Processes Machine Learning (GPML) Toolbox version 4.2, which are available at http://www.gaussianprocess. org/gpml/code/matlab/doc/manual.pdf. All data involved in the GP data processing are available at https: //data.mendeley.com/datasets/bvwfy54j2d/1.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## *Article* **The Performance and Exhaust Emissions of a Diesel Engine Fuelled with** *Calophyllum inophyllum***—Palm Biodiesel**

**Natalina Damanik 1, Hwai Chyuan Ong 2,\*, M. Mofijur 3,\*, Chong Wen Tong 2, Arridina Susan Silitonga 4,\*, Abd Halim Shamsuddin 5, Abdi Hanra Sebayang 4, Teuku Meurah Indra Mahlia 3, Chin-Tsan Wang <sup>6</sup> and Jer-Huan Jang <sup>7</sup>**


Received: 31 July 2019; Accepted: 30 August 2019; Published: 5 September 2019

**Abstract:** Nowadays, increased interest among the scientific community to explore the *Calophyllum inophyllum* as alternative fuels for diesel engines is observed. This research is about using mixed *Calophyllum inophyllum*-palm oil biodiesel production and evaluation that biodiesel in a diesel engine. The *Calophyllum inophyllum*–palm oil methyl ester (CPME) is processed using the following procedure: (1) the crude *Calophyllum inophyllum* and palm oils are mixed at the same ratio of 50:50 volume %, (2) degumming, (3) acid-catalysed esterification, (4) purification, and (5) alkaline-catalysed transesterification. The results are indeed encouraging which satisfy the international standards, CPME shows the high heating value (37.9 MJ/kg) but lower kinematic viscosity (4.50 mm2/s) due to change the fatty acid methyl ester (FAME) composition compared to *Calophyllum inophyllum* methyl ester (CIME). The average results show that the blended fuels have higher Brake Specific Fuel Consumption (BSFC) and NOx emissions, lower Brake Thermal Efficiency (BTE), along with CO and HC emissions than diesel fuel over the entire range of speeds. Among the blends, CPME5 offered better performance compared to other fuels. It can be recommended that the CPME blend has great potential as an alternative fuel because of its excellent characteristics, better performance, and less harmful emission than CIME blends.

**Keywords:** *Calophyllum inophyllum* biodiesel; palm biodiesel; engine performance; exhaust emissions; alternative fuel; transesterification

#### **1. Introduction**

Petroleum derived fuels are the main source of primary energy consumption worldwide. Because of the negative impact and limited reserve of fossil fuels, scientists have focused on the new sources of energy to replace the fossil fuel [1,2]. Renewable energy sources have been proven to create less or zero-emission energy generation and can play an important role to lower fossil fuel consumption [3]. In many countries, different types of renewable energy sources including solar, wind, hydro, geothermal, bioenergy and biofuel has been introduced [4–9]. However, some renewable energy, including wind and solar, are only available for a certain time and period and therefore energy storage is required for these kinds of sources [10]. Due to this problem, researchers attempt to find other types of energy storage material that can be commercialized [11–14]. Therefore, some scientists, especially in developing countries are more interested in the energy sources that can be kept for a long period, such as bioenergy, bioethanol, and biodiesel [15–17]. Biodiesel is one renewable energy source, which can significantly lower emissions due to fossil fuel combustion that create air pollution, global warming, and acid rain [18]. Biodiesel sources include soybean oil, sunflower oil, palm oil and cottonseed oil, *Jatropha curcas oil*, mahua (*Madhuca indica*) oil, jojoba (*Simmondsia chinensis*) oil, tobacco seeds, salmon oil, tamanu (*Calophyllum inophyllum*) oil, sea mango oil (*Cerbera odollam*), and microalgae [19–22].

Palm oil has been commonly used in Malaysia and Indonesia as a biodiesel source due to its availability and favorable characteristics [23]. The productive lifetime of palm oil is around 25 years and it has to be replanted after that period [20]. Palm oil can yield methyl ester over 80%. Since 2006, the Indonesia government has paid attention to biodiesel as part of the National Security Act of Indonesia because of world crude oil price fluctuation. It is also supported because Indonesia is the largest crude palm oil (CPO) producer. However, until 2010, the Indonesia government failed to achieve biodiesel blending targets due to the increase in the world crude palm oil price and decrease in the crude oil price. As an impact, the biodiesel price has been not competitive compared to the diesel fuel price [24]. As Ong at al. [19] reported on sensitivity analysis that differences in the price of sources will have considerable impact on the life cycle cost of biodiesel by at least 79%. However, many new policies were introduced in 2014 by the Indonesian government to promote the use of biodiesel. Ong at al. [25] suggested that a financial incentive and subsidy policy should be enforced to make the price of biodiesel competitive to diesel fuel. However, based on a cost-benefit analysis (CBA), this will enhance the net benefit of palm oil plantation and biodiesel producers but will lessen the net welfare for society and the government of Indonesia. Therefore, the policy in the future will focus on reducing costs that improve the net social benefit [24].

*Calophyllum inophyllum* seed is an inedible oil source, which has a high oil content. Therefore, *Calophyllum inophyllum* seed is also a potential feedstock for biodiesel fuel [19] in Indonesia and Malaysia due to its abundant availability. This feedstock is a biodiverse plant that was previously known as a medicinal source due to its high antioxidant content [26]. However, *Calophyllum inophyllum* is grouped into high-acid-number feedstocks that allow biodiesel production to be equipped with special treatments, such as triple-stage transesterification, degumming, and neutralization [6]. In fact, *Calophyllum inophyllum*biodiesel has a poor oxidation stability because it has about 72.65% of unsaturated fatty acids that make this fuel unfavourable for long-term storage [27]. Excessive chemical treatment for minimizing total acid number (TAN) in oil refining may lead to a reduction of antioxidant content and oxidation stability [28]. Recently, some experiments reported that the antioxidant addition into biodiesel has improved its oxidation stability.

However, recently many studies have been reported on the fractional replacement of conventional fuel by palm and CIME. There are not many studies that have been reported on the prospect of palm and *Calophyllum inophyllum* biodiesel mixture. In this work, palm and *Calophyllum inophyllum* oil were mixed prior to the biodiesel production process and compared their performance with conventional fuel in a diesel engine. This method is believed to be able today reduce the chemical process during the acid value reduction of *Calophyllum inophyllum*–palm oil compound. Moreover, the objective of this study is also to investigate the engine performance (specifically, the Brake Specific Fuel Consumption (BSFC) and Brake Thermal Efficiency (BTE) and exhaust emission characteristics NOx, HC, and CO emissions) of *Calophyllum inophyllum*–palm biodiesel mixture. It is expected that there is a potential for these blends to be commercialized in Indonesia and Malaysia due to the abundant supply of *Calophyllum inophyllum* seed oil and palm oil in these countries.

#### **2. Materials and Methods**

#### *2.1. Crude Oils*

Crude *Calophyllum inophyllum* oil and palm oils were purchased from a local store in Kuala Lumpur, Malaysia. The crude *Calophyllum inophyllum* and palm oils were mixed at 50:50 equal volume % in order to produce the CPME.

#### *2.2. Production of CPME*

Firstly, the blend was prepared by mixing 1 L of the crude oil from each source with 1 % of phosphoric acid (H3PO4, Merck Sdn. Bhd., Kuala Lumput, Malaysia) and 10 % of purified water (*v*/*v*) for 30 min. The crude oil mixture was degummed at 60 ◦C with an agitation speed of 800 rpm. The degumming process is essential to remove impurities and compounds (i.e., resins, proteins, phosphates, carbohydrates, and water residue). Next, acid-catalysed esterification was conducted. The details of the esterification process can be found in Silitonga et al. [29]. Molar ratio and catalyst percentage influence the esterification process of the oils [30]. In this study, it displayed the optimum molar ratio and H2SO4 catalyst (Merck Sdn. Bhd., Kuala Lumpur, Malaysia) concentration are 1:16.6 and 2.0 vol.%, respectively, since these parameters result in the highest esterified oil yield and fastest reaction time. According to [31], the presence of excess water can increase the formation of peroxides and increase the free fatty acid content of esterified oils. Thus, purification is crucial to remove excess water, which can be done by evaporation using a rotary evaporator, followed by the separation process with a separating funnel [32,33].

For this experiment, the esterified *Calophyllum inophyllum*–palm oil was purified by stirring the oil in a rotary evaporator (RV10 DIGITAL V IKA, Germany) at 60 ◦C with a stirring speed of 100 rpm for 30 min. The maximum pressure of the rotary evaporator was 7.2 MPa (72 bars). Following this, the esterified *Calophyllum inophyllum*–palm oil was poured into a separating funnel for the settling and left for 18 h. Karmakar et al. [21] also found that the high temperature of the purification process results in hydrolysis of the triglycerides, which in turn, removes water from the esterified oil.

Next, transesterification was done by mixing the esterified oils with 50% of methanol and 0.5 volume % of sodium hydroxide (KOH, Merck Sdn. Bhd., Kuala Lumpur, Malaysia) catalyst. The reaction mixture was stirred continuously in a jacketed reactor for 90 min maintaining the temperature at 60 ◦C. On the completion of the transesterification, the mixture was left for 4–6 h in a funnel. There are two distinct layers of liquid formed in the funnel where biodiesel was in the top and glycerol at the bottom. The glycerol was drained out from the funnel and biodiesel was washed by using sanitized water for a number of times in order to further remove impurities. The similar purification process was maintained both for the esterification and transesterification process.

#### *2.3. Production of Methyl Ester*

The CIME and palm oil methyl ester (POME) were prepared in the same manner. The crude *Calophyllum inophyllum* and palm oils were first degummed to remove impurities. The degummed oils were then esterified under the following process conditions: (1) reaction temperature; 60 ◦C, (2) stirring speed; 800 rpm, (3) reaction time; 60 min, (4) oil-to-methanol molar ratio; 1:16.6, and (5) H2SO4 catalyst concentration; 1.0 vol.%. The esterified oils were then purified to remove extraneous water present in the oils. Next, the purified *Calophyllum inophyllum* and palm oils were transesterified under the following process conditions: (1) reaction temperature; 60 ◦C, (2) stirring speed; 800 rpm, (3) reaction time; 90 min, (4) oil-to-methanol ratio; 1:8, and (5) catalyst- KOH with concentration; 0.5 vol.%. Likewise, the reaction mixtures were left to settle in separating funnels for 4–6 h after the

transesterification process. In the final step, the CIME and POME were cleaned using sanitized water several times.

#### *2.4. Characteristics of the CPME*

The characteristics (i.e., density, kinematic viscosity (KV), flash point (FP), acid value(AV), high heating value (HHV), FAME content, and oxidation stability of the CPME and its blends were examined and compared to diesel, POME, CIME, as well as their blends. The FAME content was determined by employing a gas chromatograph–mass spectrometer (Model: GCMS-QP2010 Ultra, Shimadzu, Japan) fitted with a low-bleed GC-MS column (Model: RTX-5MS, RESTEK, Tokyo, Japan) details operating condition can be found elsewhere [34]. The temperature of the flame ionization detector and split injector was 300 ◦C. The biodiesels chemical and physical properties are collected from literature as a comparison.

The FAME content in per cent (%) determined by the following Equation:

$$\text{FAME} = \frac{(\sum A) - A\_{EI}}{A\_{EI}} \times \frac{C\_{EI} \times V\_{EI}}{m} \times 100\tag{1}$$

Here, *A* is the summation of the peak areas of FAME, *AEI* is the methyl heptadecanoate peak area, which is the internal standard, *CEI* is the methyl heptadecanoate solution concentration in heptane (mg/mL), *VEI* is the methyl heptadecanoate solution volume (mL) and *m* is the methyl ester mass (mg).

The percentage (%) of the methyl ester yield can be calculated by the following Equation:

$$\text{Method } \text{ster } \, y \, \text{i} \, \text{d} = \frac{\text{F} \, \text{A} \, \text{ME} \times \, B\_{\text{c} \, \text{pmc}}}{\text{O}\_{\text{c} \, \text{p} \, \text{v}}} \times 100 \tag{2}$$

The FAME is the fatty acid methyl ester content (%), *Bcp* is the *Calophyllum inophyllum*-palm oil methyl ester weight (g) and *Oso* is the weight of the *Calophyllum inophyllum*-palm mixed oil (g).

#### *2.5. Experimental Set-Up*

Engine tests were done to study the engine performance and the characteristics of exhaust emission for CPME blends and CIME blends and the data collected compared to diesel fuel. These fuel blends were prepared in this study: (1) CPME5, (2) CPME10, (3) CIME5, and (4) CIME10. In this study, the performance parameters BSFC and BTE whereas the exhaust gases parameter NOx, HC, and CO were measured. A single-cylinder diesel engine (Yanmar YX2500CX-A 170F, Osaka, Japan) was used to investigate the performance that set in full throttle. The engine speed varied from 1400 to 2800 rpm. A BOSCH BEA 350 gas analyser was used in order to measure the emissions. The detail of the engine test-bed and emission analyser is given in Table 1.

**Table 1.** Diesel engine technical specifications.


#### *2.6. Uncertainties of the Experimental*

Generally, the uncertainties of the experiment happened due to several reasons, namely: (1) instruments type and condition, (2) instruments calibration, (3) environmental conditions, and (4) procedure of experimental. To make sure the accuracy of the data between the limit, therefore the accuracy of the experimental data should be verified. Consequently, the uncertainties percentage of selected variables, namely BSFC, BTE, CO, NOx, and HC were investigated according to the instrument's percentage uncertainties employed in the experiments. The speed accuracy, fuel consumption flowrate and time, which were ±10 rpm, ±1%, and ±0.1 s, respectively. The BSFC uncertainty was investigated by the uncertainty linearized approximation method. The details of % of uncertainties are given in Table 2.


**Table 2.** The percentage of uncertainties.

#### **3. Results and Discussion**

#### *3.1. Physicochemical Properties*

The properties for POME, CIME, CPME, and their blends are given in Table 3. It is seen that the density of the CPME (880 kg/m3) is lower than that for CIME (884 kg/m3). The KV of the CPME was found lower than that for CIME and similar to that for POME (4.4 mm2/s). In general, the KVs for CPME, CIME, and POME are inline with ASTM D6751 limit. The FP of CPME is 160 ◦C, which is above the limit of ASTM D6751 standard. The higher FP is important as it reduces the fire hazard risk, which is the main concern on fuels to handle, transport, and store [35]. However, the HHV of the CPME (37.9 MJ/kg) is found to be greater than CIME and POME (37.3 and 36.4 MJ/kg, respectively).


**3.** Comparative physicochemical properties of the fuel sample used.

**Table** 

#### *3.2. Fatty Acid Methyl Ester (FAME) Composition*

The FAME compositions of the CIME, POME, and CPME are summarized in Table 4. In general, all of these biodiesels have high palmitic acid content. However, the POME has a higher percentage of oleic acid, whereas the CPME has a higher percentage of antioxidants, such as methyl palmitic acid (C16:0), stearic acid (C18:0), linoleic acid (C18H36O2), and 9-Octadecene,1-methoxy-, (E) (C19H38O) [34]. Moreover, the CPME has a high oleic acid percentage (C18:1), with a value of 52.94 wt.%, which also serves as a lubricant.

**Table 4.** Fatty Acid Methyl Ester (FAME) composition of *Calophyllym inophyllum* Methyl ester, CIME, Palm Oil Methyl Ester (POME), and *Ceiba Pentandra* Methyl ester (CPME).


#### *3.3. Brake Specific Fuel Consumption (BSFC)*

Figure 1 shows the BSFC for diesel, CPME and CIME biodiesel blends at various engine speeds. It can be observed that all the blended fuel have higher BSFC compared to the diesel fuel except CPME5 blend. On average, biodiesel blended fuels have 16%–21% higher BSFC than diesel fuel. This finding is consistent with the literature [36–38]. Öztürk et al. [38] investigated the mixture of canola oil–hazelnut soap stock biodiesel-diesel and they found that the BSFC of blend fuel is more than the diesel fuel. The combined effects of the density, KV and HHV of the fuel caused that result [39]. During the suction stroke, biodiesel is injected on a volume basis; thus more fuels are fed inside the cylinder [40]. Consequently, more fuel is needed in order to achieve the same power because the HHV of biodiesel is lower than diesel. Among the blends, the average BSFC was highest for CIME10 blend (2.58 Ltr/kWhr) and lowest for CIME5 (2.21 Ltr/kWhr), which can be attributed by the HHV of the CIME10 blends. According to the data presented in Table 3, fuel sample CIME10 have a slightly higher heating value (43.9 MJ/kg) compared with CPME5 (43.1 MJ/kg).

**Figure 1.** Changes in Brake Specific Fuel Consumption (BSFC) of diesel, CPME, and CIME blends with speeds.

#### *3.4. Brake Thermal E*ffi*ciency (BTE)*

Figure 2 shows the BTE for all fuel samples at different speeds of the engine. It is seen that the BTEs of all fuel samples used in this study increases with the speed and maximum BTE was found for diesel fuel compared to blended fuels. This can be explained by the higher heating value and lower BSFC of diesel fuel [41]. Diesel fuel showed maximum BTE followed by the CPME5, CIME5, CPME10, and CIME10 fuels. On average blended fuel lowers 1.25%–22% BTE compared to diesel fuel. The lower viscosity and higher heating value of diesel fuel, which improves the fuel atomization; thus increased the BTEs. The data obtained from the experiment are similar to the results presented by Sharma et al. [42]. They reported that the mixed *Jatropha* and Cottonseed blend produce lower BTE than diesel fuel. The reason was explained by the poor spray formation, higher viscosity, and poor ignition quality.

**Figure 2.** Changes in Brake Thermal Efficiency (BTE) of diesel, CPME, and CIME blends with speeds.

#### *3.5. Nitrogen Oxide Emissions (NOx) Emission*

The nitrogen oxides emissions in exhaust consist of nitric oxide (NO) and nitrogen dioxide (NO2). Figure 3 shows the NOx emissions for diesel, and the CPME and CIME biodiesel blends at various engine speeds. It is evident that the NOx emissions increase with an increase in engine speed. It is clear that biodiesel blended fuels give more NOx emissions compared to diesel fuel. A similar report was found in the literature [43] for B7 and B100. The average NOx for diesel fuel was found to be 112 ppm, which is 1.5%–29% higher than the blended fuels. This can be explained by the lean air/fuel ratio because biodiesel fuel has more inherent oxygen than diesel fuel. It has been reported that oxygenated fuel blends cause higher NO*x* emissions [36]. Also, the higher KV of the biodiesel fuel leads to bigger droplets and shorter ignition delays, which affects the NOx emission [44]. In addition, the unsaturated fatty acid content of biodiesels leads to fuels higher adiabatic flame temperature than diesel fuel, which causes higher NOx emission [43].

**Figure 3.** Changes in Nitrogen Oxide (NOx) emissions of diesel, CPME, and CIME blends with speeds.

#### *3.6. Carbon Monoxide (CO) Emissions*

Figure 4 shows the CO emissions of all fuel samples at various engine speeds. The results indicate that the CO emissions are generally fewer for the biodiesel blends than the diesel fuel. Among the fuel samples, biodiesel fuel lowers 5% to 15% CO emission on average compared to the diesel fuel. The reason is described by the higher oxygen content of the biodiesels, which results in cleaner, better combustion [45,46]. CO is formed due to the incomplete combustion of the fuel due to insufficient oxygen or low gas temperature. As mentioned earlier, biodiesel fuel has a 12% higher oxygen content than diesel fuel, which accepts more carbon molecules to be burnt completely [36].

**Figure 4.** Changes in Carbon Monoxide (CO) emissions of diesel, CPME, and CIME blends with speeds.

#### *3.7. Hydrocarbon (HC) Emissions*

The comparison of emission among the fuel samples related to HC is presented in Figure 5. It was found that average HC emissions of blends were less than diesel. It is obvious that biodiesel blended fuel lowers HC emissions by 13%–22% than diesel fuel. The HC emissions can be reduced by the combustion quality improvement in biodiesel diesel blends due to the existence of excess oxygen atoms in biodiesel [47]. Similar results were reported by Mofijur et al. [37]. They explained that lower hydrocarbon emissions of moringa biodiesel-diesel occur because of higher oxygen contents of biodiesel fuel than diesel fuel. Also from the graph, it is seen that with increasing engine speeds, the HC emission decreases. Kegl et al. [48] presented similar results that both biodiesel and diesel fuels emit higher HC emissions when engines run at lower speeds.

**Figure 5.** Changes in Hydrocarbon (HC) emissions of diesel, CPME, and CIME blends with speeds.

#### **4. Conclusions**

In this study, CPME is produced by a systematic procedure that started from crude oil mixing and ended by the transesterification process. Based on the findings, the following conclusions can be made:


Finally, it can be concluded the CPME blend has potential as a diesel engine alternative fuel to lower the harmful emission.

**Author Contributions:** Conceptualization, Methodology, results and formal analysis were initiated and wrote by N.D., A.S.S. and H.C.O.; T.M.I.M., A.H.S. (Abd Halim Shamsuddin) and C.W.T. contributed to supervision. A.H.S. (Abdi Hanra Sebayang), M.M. contributed to the Mathematical derivation and results analysis; C.-T.W. and J.-H.J. checked and improved the manuscript. All authors read and approved the final manuscript.

**Funding:** This research is funded by the Centre for Advanced Modeling and Geospatial Information Systems (CAMGIS), UTS under Grants 321740.2232397 and AAIBE Chair of Renewable grant no: 201801 KETTHA. The authors would like to acknowledge the University of Malaya, Kuala Lumpur for the financial support under SATU joint research scheme (ST010-2018), the Direktorat Jenderal Penguatan Riset dan Pengembangan Kementerian Riset, Teknologi dan Pendidikan Tinggi Republik Indonesia, (Grant no. 147/SP2H/LT/ DRPM/2019) and Politeknik Negeri Medan, Medan, Indonesia.

**Acknowledgments:** The authors would like to acknowledge the University of Technology Sydney, Australia for supporting this research.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Abbreviations**


#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
