1. Introduction
Among different types of fuel cells, solid oxide fuel cells (SOFCs) result in higher efficiency in energy conversion than the others [
1]. Numerous authors have carried out investigations to model and optimize the SOFC technology. Modeling methods, generally used to minimize the challenges of SOFC development, might need different types of equipment or long run times. These models are based on the mathematical relations demonstrating the physiochemical and electrochemical reactions taking place at/between different components of a single SOFC [
2]. Moreover, developing an applicable mathematical model sometimes requires considering various variables related to the microstructure and electrochemical properties of the component-given material, SOFC working condition, and physiochemical mechanisms. Therefore, a complicated model should be solved with simplifying assumptions or complex numerical solutions. For example, a model was proposed to simulate SOFC performance at different pressures, temperatures, and gas flow by means of computation three-dimensional fluid dynamics (CFD) simulation [
3]. Additionally, a three-dimensional model was proposed to investigate the maximum temperature and distribution at different working states [
4]. Moreover, a two-dimensional thermos-fluid model for analyzing the effect of gas composition, flow rate, and porosity as a microstructural feature was introduced by coupling an electrochemical model with a CFD model. So, this model can describe transport and reaction phenomena in a SOFC [
5].
Fuel cell performance is normally characterized by its polarization curve, which is a plot of its potential versus current density. The obtained data can be further used to predict the current-voltage dependency by Machine Learning (ML) models. These predictions are usually of higher accuracy than conventional modeling results, and the algorithms within the ML can establish authentic relationships based on empirical data. Supervised learning has increasingly attracted attentions in the field of material science and engineering thanks to its unique advantages compared with other mathematical and modeling approaches [
6,
7,
8,
9]. The major difference between ML and other mathematical modeling relies on the fact that statistical optimization requires established relationships among variables and some initial assumptions. In such a context, ML models could solely find the pattern and relation between all the data without any presumption [
10]. The major goal of ML is to find a hypothetical function enabling scientists to predict new output values based on the proposed model. It can also save time, cost, experimental materials, etc. [
11].
Traditionally, the commercialization process of the SOFC system includes several factors such as materials consumption, numerous experiments, and long-term stability testing. Considering these factors is time-consuming and requires a great amount of funding. Regarding hindering these limitations, some mathematical models and simulations were conducted, as was mentioned earlier. One of the approaches within which the mathematical models were utilized is the optimization of the cell parameters. Nevertheless, complexity of the SOFC system and calculations, simplifying assumptions, and the prerequisite knowledge over the physiochemical characteristics of a cell lead to insufficient accuracy of the predictive models [
12]. Machine learning methods could be exploited to overcome the aforementioned limitations, and also a comparison of these methods’ prediction results would provide the best possible model with higher accuracy [
13].
Considering the ML method as a new and proper modeling candidate, investigations have been focused on this area. Some researchers conduct ML methods for performance prediction, and some focus on comparing different models to enhance prediction results with less errors. In this regard, an artificial neural network (ANN) was conducted to predict the output voltage of cells in anode support SOFC, and some variables such as temperature, porosity, cathode layer thickness, electrolyte thickness, anode thickness, and current density were selected as input features. The current–voltage response was predicted under different circumstances, and the standard error of prediction of 1.705% was reported [
14]. In another research, for a data set based on different parameters of Proton Exchange Membrane fuel cell (PEMFC) and its polarization behavior Support Vector Machine (SVM) and ANN techniques were utilized for comparing these methods and cell performance prediction [
15]. Moreover, in another study on PEMFC, 9725 experimental data were collected from an electric bicycle powered by PEMFC, and these data sets were modeled by SVM and ANN. The predicted voltage–current, power–current, and efficiency power were achieved with 99% of accuracy for SVM and 97% for ANN [
16]. Furthermore, support vector regression (SVR) and random forest (RF) algorithms were conducted to predict the performance of a SOFC cell, and the effects of hydrogen purity on the fuel mixture and temperature on cell output were investigated by ML. The results showed that the SVR algorithm depicts lower error values and better prediction [
17]. Moreover, 858 data from 30 pieces of SOFC stack tested at four different operating conditions were utilized to predict stack performance using back propagation neural network (BP), SVM, and RF algorithms. These algorithms showed prediction errors of 1%, 4%, and 6%, respectively [
18].
Various experiments, modeling, and data-driven methods have been applied and used to capture the relations between the cell parameters of SOFCs and current–voltage dependency [
19,
20,
21,
22]. Previous studies built some machine learning models and predicted target values or compared the models and described the best one [
16,
17,
18]. In this study, a group of data related to the architectural and operational variables was used. Convolutional machine learning models and a multilayer perceptron (MLP) neural network were implemented to predict the current–voltage dependency as a function of the operational temperature, anode support thickness and porosity, electrolyte and cathode functional layer thickness, and compare the models’ efficiency to experimental results. Moreover, a comparison was performed between the obtained machine learning model and previous ones which employed neural networks and mathematically developed equations to justify the accuracy of predicted values from previous experimental results [
14,
23]. Then, the relations between cell design, operating temperature, and output voltage of a single cell of an SOFC were captured, and the performance was evaluated.
4. Results and Discussion
In order to measure the prediction score of the regression models, the R
2 score (0 < R
2 < 1) was used. The higher the value of R
2 indicates a more accurate prediction for the model, and the lower value means a reduced ability of the model. All the models were trained with a training set as mentioned in the previous section and the R
2 score was measured. To evaluate the models’ errors, some error metrics such as mean absolute error and mean squared error were used. A comprehensive evaluation of all models was performed, and the result is represented in
Table 3.
The value of R
2 and errors represented in
Table 3 for different models showed that the best machine learning model for predicting the relations between the dataset variables is a multilayer perceptron (MLP). Multilayer perceptron model has the highest R
2 and the lowest mean squared and absolute error. This model with perfectly tuned hyperparameters had good generalization. On the other hand, the lowest R
2 and the highest mean squared and absolute errors are related to the linear regression model. These scores could be expected because, as mentioned before, the correlation value close to 1 or −1 could be interpreted as a linear relationship between two variables. In contrast, the correlation value close to 0 could be construed as non-linear relation between two variables. By considering the relations between the features and the target (
Figure 5), it can be interpreted that the relations between the features and the target are mostly non-linear since the essential components (after using PCA, the features mapped onto new components which carry same information for model) have a correlation value of −0.54, −0.45, 0.44, and −0.36 with the target, respectively. Therefore, it was observed that all the features have non- and semi-linear relation with the target value. It was found that linear models consider a linear function to predict the targets. So, it seemed that the linear model could not learn the relations properly. In the neural network case, with perfect hyperparameters tuning, the model could learn the relations excellently. However, it usually works well on high-volume data. The other models, i.e., kernelized models (support vector regressor (SVR) and Gaussian process(GP)), decision tree and its ensembles (RF and GB), and K-nearest neighbors predict the hidden function of the data reasonably. Nevertheless, the decision tree is an unstable model with high bias or variance having the highest error among the other models. This model seemed to be underfitted or overfitted. Ensembles of the decision trees may reduce the errors and improve the R
2 score.
Prediction error plots were used to measure the ability of the models in prediction and generalization, as shown in
Figure 8. The higher prediction accuracy is related to the graph in which the predicted values are closer to 45° line. It other words, the dispersion between the predicted and test target data is small, and these two values are close. The highest R
2 score is related to the MLP neural network, which exhibits the best generalization. The highest dispersion is related to the linear regression model, and then to the decision tree (
Figure S1e–h). Other models showed low dispersion in prediction error plots. The Gaussian process regressor is the best model after MLP neural network due to its flexibility in different complexity and the strong potential to find complex relations between the data. The decision tree model exhibits the highest dispersion, which means the lowest R
2 score after linear regression. Low generalization of the model can be seen in the prediction error visually. Random forest and gradient boosted trees exhibit better dispersion close to 45° line compared to the decision tree. The key factor in the experimental dataset is the bias during the experiments, such as human error, experiment error, or calculation error. Therefore, the bias of the model can be reduced by using the gradient boosting model. This correction is due to modifying a series of decision trees in which each tree tries to modify the previous one. So, in this dataset, the probability of bias is high since the data was collected from an experiment. Hence, a gradient boosting model could be more efficient than a random forest model, which was also proved by the corresponded R
2. In the case of K-Nearest neighbors, this is a classic and non-parametric model, and it could be either linear or non-linear based on data, and this R
2 score is not unexpected.
It is worthy to note that the number of samples can significantly affect models’ accuracy and generalization. The more samples, the better the accuracy score and generalization are expected. In some cases, however, a higher number of samples is helpful for a model, though the model then may require more powerful resources such as Central processing unit (CPU), memory, and graphics processing unit (GPU). Hence, learning curve plots were used to measure the effect of the number of samples on the trained machine learning models, and the results are shown in
Figure 9. The lines are the mean score value, and the shaded area around each lines indicate the variance of the model.
According to
Figure 9, the linear model score was not influenced by the number of samples. It is reasonable because, with an increase in the number of samples, the relationship between variables defiantly did not tend to be linear. Therefore, it could be concluded that the complexity of the problem is higher than the linear model complexity. In all other cases, increasing the number of samples results in an improvement in models’ scores (
Figure S2e–h). In other words, increasing the number of samples could provide ample information space for the model because each new sample contained new information. Therefore, for the appropriate model, further data supplementation in this system could be helpful until the cross-validation line reaches a stable state. After the semi-horizontal trend in the cross-validation line, increasing the number of samples is not reasonable because the extensive data needs more robust resources and increased calculation expenses. It should be mentioned that the linear algorithms in the linear regressor and SVR led to a lower cross-validation score than other non-linear models because of the non-linear relations between features and target value, which were discussed earlier. Hence, it can be concluded that the number of samples was enough to build the general models. These results, therefore, approved the reliability of built general models in this study.
During the increase in the sample amount, the variance of the models’ prediction became smaller. It is due to the information that the samples carried and gave the generalization possibility to these models. In the first step of the learning curve, the two best models, the Gaussian process, and MLP neural network start with high variance. During the training phase, the variance gradually decreased; in the final step, the variance reached the smallest value. The neural network model could predict the targets with about 250 training samples in this case. This model captured all the relations in just about 250 training samples. The model accuracy did not improve during the increase in the training samples.
The main advantage of ML over utilizing mathematical modeling is its convenience and accuracy. A mathematical model proposed by AI. Milewski [
23] predicted the behavior of an anode-supported cell from experimental data presented by Zhao and Virkar in 2005 [
24]. Equation (6) was introduced by this model to represent the voltage–current dependency in a variety of physical and operational conditions.
where
is maximum voltage,
,
,
, and
are fuel utilization factor, maximum current density, area specific internal ionic resistance, and area specific internal electronic resistance, respectively. Combining laws and relations, such as the electrical laws, solid material properties, and some electrochemical laws, must be considered to drive this kind of model and the combination requires assumptions and a wealth of knowledge. Therefore, these mathematical models are not flawless in prediction, and it is challenging to propose a mathematical model regardless of assumptions. For example, the model proposed by Ref. [
23] assumed the independency of temperature on the electrical resistivity, which causes a difference between the mathematical model predictions and experimental results. However, utilizing the ML method with different algorithms can increase accuracy and reduce calculation costs. The MLP model proposed in this paper predicts the effect of temperature, anode porosity, electrolyte, and CFL thickness on the current–voltage dependency with 1%, less than 1%, 1%, and less than 1% percentage error values, respectively. The resulting plots are shown in
Figure 10a–d. The changes imposed by variation in the abovementioned parameters showed reasonable compatibility with the literature. Briefly, the elevation of temperature results in increasing the output voltage at a constant current due to an increase in Ni/YSZ conductivity [
40]. Additionally, the anode porosity plays an effective role on output voltage in the same situation. By increasing the porosity to an optimized value, the output voltage increases because of the gas diffusion facilitation and proper distribution of the active surface area [
41,
42].
According to
Figure 10, the predicted results followed the experimental data properly. The relative error values proposed by ref [
33] for the temperature, anode porosity, electrolyte, and anode thickness were in the range of 2 to 7%. It is worth mentioning that the ANN prediction for this dataset in ref [
12] illustrated an imprecise prediction line for 32% porosity within the anode while according to
Figure 10, the MLP model in this paper showed a more accurate prediction.
5. Conclusions
Machine learning method has shown numerous benefits, such as saving time and energy within the long-term experimental procedures and being cost-effective, making it an appealing approach for prediction in various study fields. Therefore, this method is a viable approach to evaluating solid oxide fuel cells due to their complex fabrication process and operational conditions. Regarding this paper, eight machine learning models were conducted to compare their accuracy for predicting the output voltage of an anode-supported solid oxide fuel cell, including linear regressor, K- nearest neighbors regressor, support vector regressor, random forest regressor, gradient boosting regressor, Gaussian process regressor, and multilayer perceptron regressor. The latter was implemented with two hidden layers, Relu activation function, and 300 neurons for each hidden layer. The results were discussed according to three metrics, mean absolute error, mean squared error, and R2 score, which evaluated both the models’ error range and accuracy. It was indicated by the results that the complex models such as multilayer perceptron and Gaussian process regressor provided higher accuracy due to the non-linear correlation between the features and target values; on the other hand, linear and support vector regressors could not perform an efficient prediction because of their linear solver and low complexity. The multilayer perceptron regressor showed the highest R2 score among the aforementioned models with a 0.998 R2 score, and the lowest mean absolute and squared error of 0.006 (V) and 9.6 × 10−5, respectively. The Gaussian process regressor yields an R2 score of 0.996, mean absolute, and squared error of 6 × 10−3 (V) and 10−4, respectively. Therefore, regarding these observations, the MLP regressor is a robust model able to predict the output voltage of an anode-supported solid oxide fuel cell based on its operational temperature, anode-supported porosity and thickness, electrolyte and cathode functional layer thickness, and current density. Afterward, the MLP method was conducted to predict the interrelations between features and current–voltage dependency, and the results indicated that the model prediction lines followed the experimental data prosperously and these prediction lines are more accurate than previous studies.