ANN’s most common technique, known as feed forward MLP artificial neural network (MLP-FF-ANN), is one of the techniques that relies on memory less and thus the reflection in response to the result is much faster than the feedback type [13]. In this study, MLP-FF-ANN generated by JMP version 10 (SAS, USA) was used for simulation because of its flexibility and ease of use when compared to the Statsoft Statistica Release 7.0 software [14]. This network type consists of a system of multi-level interconnected “neurons” or nodes that are formed in three layers, namely input layers (independent variables), hidden layers (one or more), and output layers (dependent variables) [15, 16]. The input variables were calcination temperature (°C), based ratio (%) and catalyst dosage (g), while the output was CO₂ conversion (%). The parameters were chosen based on our previous study [11, 17], as they could give great influence on the performance and physicochemical properties of the prepared catalyst. The model was trained by using the Levenberg Marquardt Learning Algorithm [14, 18]. The respected R² and RMSE were computed using Eq. (1) and (2), respectively:

The predicted value is signified by where t_i the measured value of the experimental samples is, td_i signifies the predicted value, and N represents the total of samples. The range calculation for R² is in between 0.0–1.0. The lowermost value of R² is considered weak, whilst the highest value (near to 1.0) indicates suitability to be chosen as the best predictor [18]. Next, the efficiency model (EM) by Nash and Sutcliffe [19] was used to test the fit between measured and modelled data using the Eq. (3):

Where, E refers to coefficient of efficiency, O_m represents the measured value, Q_p indicates the predicted value and Q_{mea_avg} denotes an arithmetic average measured value. If modelled values perfectly match the measured ones, then E is equal to 1.

XRD analysis was performed to observe the presence of active oxide phases on the prepared RMS (5:35:60) catalyst in different calcination temperatures of 400, 700, 900, 1,000 and 1,100°C. Calcination temperature would influence both the crystal phase and crystallite size of the catalyst. Some processes including change in structure, arrangement of crystal lattice and active phase formation could also occur during sintering [11, 20]. The XRD diffractograms of RMS (5:35:60) catalysts calcined at different temperatures as shown in Fig. 1.

As shown in Fig. 1, the phases for RMS (5:35:60) catalysts were highly amorphous and most of them were dominated by alumina support, Al₂O₃ at 2θ values of 66.86°, 45.79°, 37.73°, and 39.42°. The amorphous phases may be due to the presence of the lanthanide element, samarium in the catalyst and may have hindered the crystallinity of the catalyst phase. RMS (5:35:60) catalyst heated at 400°C and 700°C shows similar diffractograms with the presence of alumina cubic peaks only. Meanwhile, RMS (5:35:60) catalysts heated at 900 to 1,100°C showed the existence of tetragonal RuO₂ at 2θ values of 28.08°, 35.09°, and 54.53° besides cubic Sm₂O₃ peaks at 27.88°, 32.31°, and 45.36° in addition to the cubic of Al₂O₃. Interestingly, the presence of new small peaks of MnO₂ with tetragonal phase at 2θ values of 36.99°, 56.40°, and 42.82° were observed for the catalyst calcined at 1,000 and 1,100°C. However, the appearance of low MnO₂ intensity peak in the diffractogram was in accordance with research conducted by Xu et al. [21] who also observed a broad peak of MnO₂ which indicated the amorphous nature of the product. Therefore, 900°C of calcination temperature is not sufficient to increase the crystallinity and size of the MnO₂ species thus, it was not observable in the diffractogram. Besides that, a few peaks on the diffractogram that overlapped each other were also observed. These overlapping peaks occurred due to very similar crystallographic characters from the PDF XRD profile.

To determine the oxidation state of elements in the catalyst, the relative contents of each state over the catalyst surface were deconvulated using XPS software based on the binding energy. The relative surface charge was adjusted against the C 1s line (284.50 eV) as a standard. Fig. 2 shows the deconvulation of XPS spectra for O 1s, Al 2p, Mn 2p and Sm 3d. Oxygen constitutes a high percentage in the surface composition, which was 25.42%. The O 1s photoemission spectrum exhibited single-lobed envelopes. The deconvulation produced two peaks, which suggests the presence of two different elements. The strongest peak had an E_b value of 531.50 eV and was assigned to Al₂O₃ or oxygen of the Sm³⁺-O groups. The presence of a weak peak at an E_b value of 529.96 eV was assigned as oxygen of Sm²⁺-O groups. This is related to the oxygen in the crystal lattice of Sm₂O₃ [22]. Aluminium constituted a high percentage in the surface composition, which was 65.04%, confirming the Al₂O₃ as a support. Deconvulation of Al 2p produced two peaks that indicated Al³⁺ in different chemical environments. The E_b values of 73.40 and 74.47 eV, were credited to the presence of Al³⁺ in Al₂O₃ and Al(OH)₃, respectively. Meanwhile, Mn 2p consisted of two doublets which were Mn 2p_1/2 and 2p_3/2 which are equal to 11.5 eV. In this study, the high binding energy in the range of 642 – 647 eV was ascribed to the presence of Mn⁴⁺ from the MnO₂ species [23] and was aligned with the observation obtained from the XRD analysis in Fig. 1.

The peak of samarium formed two doublet structures of 3d5/2 and 3d3/2 equal to 28 eV as a result of spin-orbit splitting and their satellites in rare-earth compounds as also explained by Chung et al. [24]. The first doublet corresponded to Sm 3d5/2 at a binding energy of 1,082.33 eV that was attributed to Sm3+ ion under its oxidized form in metal oxide, Sm2O3. The observation of satellite Sm2O3 shouldered at 1,085.73 eV was associated with a strong charge-transfer effect, which leads to an increase in the number of unpaired 4f electrons by more than one of the samarium oxides [25]. It was also reported that Sm was in a trivalent state when adsorbed onto transitional metals due to the close binding energy at 1,082.33 and 1,085.73 eV [26, 27]. This is supported with XRD where Sm₂O₃ was observed in the diffractograms.

The ESR spectra of the RMS (5:35:60) catalyst calcined at various temperatures are shown in Fig. 3. From the spectra, all of the catalysts showed a g value of 2.12 with different intensities. The shift from the value was caused by the electron spin-orbit interaction in atomic orbital. The ‘wing’ that forms from both sides at the g value of 2.12 was attributed to Mn⁴⁺ as suggested by Chakradhar et al. [28]. This result correlates well with the result obtained from the XRD and XPS analysis in Fig. 1 and 2 whereby the manganese species that exists in this catalyst were MnO₂ (Mn⁴⁺). This result is similar to Phan et al. [29] who identified that MnO₂ will exhibit a symmetrical center which is caused by a paramagnetic property at a magnetic field about 300 mT. From the ESR spectra, at calcination temperatures of 1,000°C and 1,100°C, both peaks overlapped and exhibited the highest peak intensity as compared to 900°C. It can be clearly seen in XRD diffractograms that the presence of MnO₂ peaks in the catalyst calcined at 1000 and above. However, the peak of MnO₂ was not detected in XRD analysis calcined at 900°C which can be explained by the lower signal ESR peak due to lower intensity and crystallinity of the MnO₂ species. Therefore, the paramagnetic properties of the catalysts were said to be higher at 1,000°C and 1,100°C which could contribute to the maximum catalytic activity of CO₂ conversion.

A H₂-TPR analysis was performed to provide information on the reducibility of various chemical species contained in the catalyst, the degree of interaction between metal-support, and to determine the reduction temperature. The H₂-TPR profile of RMS (5:35:60) was depicted in Fig. 4. The H₂ consumption of the profile provides a cursory measure level of coordination and availability of the surface capping oxygen ions in the catalyst [30].

At calcination temperatures of 900, 1,000 and 1,100°C, four reduction peaks were detected at 268.2, 433.5, 594.2, and 976.7°C, respectively. The first reduction peak at 268.2°C was owing to the reduction temperature of RuO₂ reduction to metal Ru where the H₂ consumption was 0.12 cm³/g STP [31]. It can be observed that only a tiny amount of H₂ was consumed on the catalyst. As observed by Peluso et al. [32], the second reduction peak at 433.5°C of H₂ consumption, 0.02 cm³/g STP belonged to the reduction of Mn⁴⁺ to Mn²⁺. On the other hand, the reduction peak at 594.2°C was due to the reduction of the Mn-O-Al species with a total H₂ consumption, of 0.02 cm³/g STP [33]. The last reduction peak at 976.7°C was consigned as the bulk oxygen reduction from SmO₂ to Sm₂O₃ where the amount of H₂ consumption was recorded as 0.13 cm³/g STP. The reduction peaks at 1,100°C showed a shift to higher temperatures, implying stronger interactions between active species and alumina support which led to a decrease in reducibility. The outflow of H₂ to the catalyst surface increased the H₂ adsorption area and improved the hydrogenation activity. The H₂-TPR profile demonstrated that γ-Al₂O₃ was also an inert carrier, so no reduction peak of γ-Al₂O₃ was observed [34].

A CO₂-TPD analysis was done to study the basicity of the catalyst surface and determine the CO₂ and H₂ sorption profiles on its surface. Lin et al. [35] found that supported noble metal catalysts have strong metal-support interactions, which affects the performance of active species, and favors the adsorption through electron-deficient state of metal nanoparticles. The difference in desorption profile was caused by the coordination of CO₂ to metal oxides with different energies. The CO₂-TPD profiles were depicted in Fig. 5.

The peak at the lowest temperature of 117.6°C was due to desorption of low chemisorption or physical adsorption of CO₂ on the catalyst surface [36]. Nevertheless, the peaks at 210.6°C and 279.4°C could be ascribed to the inorganic carboxylate desorbed on the surface of the catalyst as suggested by Cox [37]. In addition, the TPD profile’s peak area of CO₂ desorbed from the surface is used to determine the catalyst’s total basicity. Therefore, the broad peak areas at 210.6°C and 279.4°C indicated higher basicity surface of catalysts due to many CO₂ gas desorbed on the basic sites of the catalyst. Aside from that, the amount of CO₂ adsorbed on the catalyst at 1,000°C was also the highest at 210.6°C and 279.4°C with 48.62 cm³/g and 34.00 cm³/g at STP respectively. As a result, the catalytic activity of the catalyst will be higher and more active at reaction temperatures above 200°C.

FESEM analysis was used to study the morphology of the RMS (5:35:60) catalyst supported on alumina. Fig. 6 displays the FESEM micrographs at 50000X magnification of the RMS (5:35:60) catalyst calcined at 400, 700, 900, 1,000 and 1,100°C for 5 h. The micrograph for different calcination temperatures showed no significant changes of morphology. As can be observed, the particles became more agglomerated with increasing calcination temperatures. The micrograph also showed that when the calcination temperature was raised to 1,000°C, the catalyst had more pores and this structure could help in the CO₂ methanation activity. From the morphology of the catalyst, it revealed more agglomeration occurred by providing higher surface areas for the active species to absorb the gasses which contributed to higher catalytic activity. At a calcination temperature of 400°C, the catalyst surface was tightly packed, which results in a small surface area, which limits the active sites of the methanation reaction and thus reduces the catalytic activity performance. At a calcination temperature of 1,000°C, micrographs showed that small spherical particles were agglomerated on the catalyst surface which can provide the highest catalytic activity.

The composition of the element that was present on the catalyst surface was confirmed by EDX analysis. The EDX analysis revealed a low atomic ratio of Sm, Mn, and Ru with below than 1.00%. This is most likely due to the absorption of alkaline metal, Sm, Ru and Mn in highly porous and high surface area alumina beads. This is in agreement with the FESEM analysis which showed many pores on the surface. However, when the calcination temperature was raised to 1,000°C, the atomic ratio of each element increased which are Sm (0.72%), Mn (0.79%), except for Ru (0.37%). This is supported by the XRD where the peak of the RuO₂ species was very small in intensity. Previous research conducted by Zielinska et al., reported that the content of elements increased with the increasing calcination temperature [38]. However, when the catalyst was calcined at 1100°C, the element composition of Sm and Mn were slightly decreased with 0.49% and 0.64%, respectively except for the Ru element which showed an increase with 0.54%. This observation was corroborated by XRD analysis and again showed the appearance of the Mn and Ru in the catalyst samples.

The pore structure and surface area analyses were measured by nitrogen adsorption analysis with the Brunauer–Emmett–Teller (BET) method. It was found that the BET surface area of the RMS (5:35:60) catalysts decreased with each increment of calcination temperature. The reduction of surface area among the catalysts might be due to agglomeration of metal oxide particles as shown by the FESEM analysis. At a 400°C calcination temperature, the BET surface area was higher by 155.03 m²/g with a 89.66 nm average pore diameter, indicating a smaller particle size. This was verified by a H₂-TPR analysis (Fig. 3) in which the H₂ consumption showed a reduction at lower reduction temperature due to an increase in the particle dispersion [39]. When the calcination temperature was increased until 1000°C, the BET surface area showed only 41.36 m²/g with an average pore diameter of 207.20 nm. Increasing the calcination temperature increases the pore size. This is due to the fact that large pores can lead to sintering, resulting in a decrease in pore volume and the formation of dense solid. Although, higher calcination temperatures reduce BET surface area, but have the advantage of ensuring the presence of a stable structure on the catalyst.

The screening of monometallic and bimetallic catalysts has been discussed in our previous study [40]. The lower CO₂ conversion for monometallic and bimetallic oxides has steered to the trimetallic oxide catalyst. However, the higher reaction temperature was not compatible with the requirement of industry to achieve a low reaction temperature. Thus, the effect of calcination temperatures on the potential catalyst of RMS (5:35:60) with respect to the catalytic performance using FTIR analysis was investigated. At 400, 700, 800 and 900°C, the CO₂ conversion were 94, 96, 89 and 91%, respectively. Meanwhile, at 1000°C calcination temperature, the CO₂ conversion showed a 100% CO₂ conversion, followed by 1,100°C, with a 98% which was slightly decreased at a maximum 400°C of reaction temperature. From the diffractograms in Fig. 1, the noticeable peaks of MnO₂, RuO₂ and Sm₂O₃ showed that these species contributed in determining higher catalytic activity (100%) of RMS (5:35:60) calcined at 1,000°C.

The next parameter step was to optimise the RMS (5:35:60) catalyst for various samarium loadings to observe its effect on the performance of catalyst. The ratios of 50–85 wt% of samarium oxide catalysts were selected. Low performance (< 30%) was observed for all the prepared catalysts (50–85 wt%) at 200°C reaction temperature except for the 60 wt% based loading which able to convert 65% of CO₂ at the same reaction temperature. It was clearly showed that an appropriate amount of samarium could increase the activity of the catalyst. As RMS (5:35:60) catalyst gives highest CO₂ conversion at 200 and 400°C reaction temperature (100%), it can be concluded that 60 wt% is the optimum loading for the samarium content. When the Sm content was increased beyond the optimum loading (60 wt%), the CO₂ conversion decreased. The excessive loading of Sm may block the active species on the catalyst surface as reported by He et al. [41]. Therefore, the trend for based loading is as follows: 50 < 55 < 85 < 80 < 75 < 70 < 65 < 60 wt%.

The parameters for various catalyst dosages were also investigated to find an optimum value for CO₂ conversion. The CO₂ conversion began to increase drastically at a reaction temperature of 200°C for each dosage. At a reaction temperature of 400°C with 3 g of catalyst or more, a complete CO₂ conversion (100%) was observed. Almost complete conversion of 98.53% occurred for 5 g of catalysts. This observation is mainly due to the increase in the number of active sites on the catalyst surface which are affected by the increase in the catalyst dosage. This finding was consistent with Su et al. [42], who found that when the catalyst dosage is increased, it would also increase the catalytic activity. Therefore, the catalyst dosage trend for CO₂ conversion was 3 < 5 < 7 g.

In this study, a neural network technique was applied to develop the best CO₂ conversion prediction model using three hidden nodes. The ANN model was trained using the Levenberg Marquardt Learning Algorithm. The factors were calcination temperature, °C (900–1,100°C), based ratio, % (60–80%), and catalyst dosage, g (3–7 g). The parameters were selected based on preliminary results. Table 1 shows the experimental and predicted values for CO₂ conversion.

In general, the percentages of CO₂ conversion obtained were high. The efficiency test of modelling was also was performed to validate the quality of the model between predicted and measured values. Based on the findings, almost all prediction values for CO₂ conversion were near to 1. According to Nash and Sutcliffe [19] and Grunwald and Frede [44], the closer the model efficiency is to 1, the more accurate the model. Table 2 summarizes the statistical parameters of the selected ANNs: Number of hidden nodes, R², RMSE and sum frequency. Selection of the network for prediction of CO₂ conversion was carried out with a trial-and-error technique. The optimal number of hidden nodes is three. The training and validation of R² and RMSE values are 0.99/0.32 and 0.99/0.22, respectively.

Then, the equation for CO₂ conversion calculation was developed based on the ANN analysis. The equation can be expressed as in Eq. (4)–(6):

Firstly, the calculation of three hidden nodes was performed:

Lastly, the prediction model of CO₂ conversion was calculated using Eq. (7):

where, H1, H2 and H3 represent the hidden nodes.