Applications of soft computing techniques for prediction of pollutant removal by environmentally friendly adsorbents (case study: the nitrate adsorption on modified hydrochar)

Due to the ongoing growth of the global population and agricultural activities, the consumption of fertilizers that include nitrogen has significantly expanded in recent years (Moradzadeh et al. 2014; Nguyen et al. 2021b). However, excessive fertilizer use is associated with the increased of the nitrate concentrations in the water bodies, causing serious environmental problems (Valiente et al. 2020). The negatively charged nitrate ions cannot bind with soil particles and are released from agricultural land drainage into main water supplies (Hafshejani et al. 2016a; Goeller et al. 2019). The increased nitrate concentration causes eutrophication and adverse health effects on pregnant women and infants (Long et al. 2019). Therefore, nitrate must be removed from agricultural runoff before entering natural water bodies.

The removal of nitrate has been studied using several techniques including chemical denitrification, biological denitrification, ion exchange, reverse osmosis, and adsorption (Bhatnagar & Sillanpää 2011). Adsorption is one of them and has been strongly advised as a remediation technique that employs affordable technology and has great decontamination qualities (Divband Hafshejani & Naseri 2020; Salam et al. 2020). Such variables such as initial solution pH, adsorbent amount, adsorbate concentration, contact time and presence of competing ions. Significantly affect adsorption capacity and removal efficiency. Therefore, mathematical and experimental models are required to interpret the interactive effects of these variables and predict adsorption performance. Conventional process optimization and data collection can be time-consuming and costly. In this regard, adsorption performance may be enhanced by utilizing advanced tools like artificial intelligence, which has a great capacity for solving complex issues.

One of the artificial intelligence models is the artificial neural networks (ANN) that is inspired by the structure of the human brain. Each ANN consists of three types of layers: input, hidden, and output layers. The number of neurons in the input layer is equal to the number of input variables, and the number of neurons in the output layer is equal to the number of output variables (Zhu et al. 2019). The input and output layers act as independent and dependent variables, respectively. In adsorption studies, the input and output layers are the independent variables and adsorption capacity, respectively. The hidden layers are the main part of neural network processing, which can include several different layers and neurons. The number of nodes, the number of hidden neurons, and subsequently the best network structure with the lowest deviations is achieved based on the trial and error method (Ding et al. 2021; Wei et al. 2021).

A support vector machine (SVM) is a common forecasting technique in data mining. This method discovers the patterns in the data and makes predictions using those patterns. The support vector machine makes its predictions using linear and non-linear combinations on a set of training data called support vectors. The construction process consists of two steps training and testing. In the training stage, the model learns the relationships between the data set and it can then estimate new data from the detected pattern. Gene expression programming (GEP) is also one of the soft computing techniques and a branch of evolutionary algorithms based on Darwin's theory of evolution. This algorithm has a high ability in modeling non-linear and dynamic processes (Amar et al. 2022). Artificial intelligence models are trained to perform using existing laboratory data. When the difference between laboratory and computational values by the model is acceptable, the learning process is achieved (Ding et al. 2021; Maurya et al. 2022). Research has shown that artificial intelligence models successfully predict the ability of different adsorbents to remove contaminants from water sources. Zhu et al. (2019), used artificial intelligence models (artificial neural network (ANN) and random forest (RF)) to simulate the adsorption of zinc, lead, cadmium, nickel, arsenic, and copper on 44 types of biochar. The models used to predict the adsorption capacity were trained and optimized according to the characteristics of biochar, temperature, pH and initial concentration of pollutants. The results showed that the random forest model (R² = 0.973) is more accurate for predicting absorption performance than the artificial neural network model (R² = 0.948). The researchers believed that the use of artificial intelligence models in predicting the ability of adsorbents can significantly reduce the workload of testing. Wang et al. (2021), predicted the ability of graphitic-C₃N₄ to absorb lead, cadmium, and mercury using artificial intelligence models (deep neural network and transfer learning). The results showed that these models are well able to evaluate the adsorbent's ability to remove the studied heavy metals well with only one-tenth of the experimental data (Mahmoud et al. 2019). Adsorption of phosphate on nanoscale zero-valent iron was simulated using artificial neural network models. The results showed that the artificial neural network with the structure of 5-7-1 predicted phosphate removal well (R²: 97.6%). The results of sensitivity analysis showed that among the five input variables to the model (phosphate concentration, adsorbent dose, stirring rate, initial pH reaction time), pH was the most influential input. Ullah et al. (2020a), predicted that the adsorption of lead ion on rice husk using artificial intelligence modeling (feed forward back-propagation neural network (FFBPNN) and Levenberg–Marquardt (L–M) training algorithm). The results showed that the artificial neural network model used has high reliability in predicting the absorption capacity of rice husk for lead removal. Ullah et al. (2020b), developed artificial intelligence models to predict the Zn (II) absorption of biochar derived from rambutan (Nephelium lappaceum) peel. The results showed that all models had high efficiency, but the ANFIS model had the best performance with an accuracy of 90.24% compared to the ANN (88.27%) and MLR (59.14%) models. Chakraborty & Das (2020), established an ANN model to predict the adsorption efficiency of chromium (VI) on nanocomposite sawdust biochar. The ANN model helped them to develop the appropriate adsorption mechanism and the best possible equation for adsorption of chromium (VI) by modified biochar. Nguyen et al. (2021a), showed that the GEP model can successfully describe the process of cesium adsorption by Prussian blue as a function of adsorbent particle size, pH, contact time, initial dye concentration, and temperature. This study looked into the ideal circumstances for removing nitrate from aqueous solutions. Then, with the purpose of minimizing expensive, time-consuming, dangerous, and laborious laboratory work, several models of soft computing techniques such as artificial neural network (ANN) and support vector machine (SVM), and gene expression programming (GEP) were applied for describing the relationship between the affecting factors on nitrate adsorption by hydrochar modified by magnesium chloride and adsorption capacity. As mentioned, various studies have shown the successful applications of artificial intelligence models for absorbing various pollutants, but the type of pollutants and adsorbents affects the efficiency of the mentioned models which needs to be discussed. In other words, according to the type and amount of speed input variables, the prediction accuracy of each of these methods is different, which should be evaluated and the most successful model should be selected according to the conditions of the problem.

The sugarcane bagasse (SB) was provided by a local farm in Khuzestan, Iran. To remove dust and other contaminants, the sample was washed multiple times using tap water, and then distilled water. The washed samples were dried at 80 °C before incorporation into the synthesis processes. Ten grams of sugarcane bagasse was mixed with deionized water (60 mL) in a stainless-steel autoclave (100 mL maximum capacity). The mixture was hydrothermally treated at a fixed temperature of 220 °C for 4 h to prepare the SB hydrochar. The autoclave was then cooled to room temperature. The hydrochar was collected and washed with deionized water until the pH was lowered to a constant value. The resulting hydrochar was dried in an oven for 24 h at 50 °C and the SB hydrochar was labeled HCSB (Silva et al. 2017; Bento et al. 2019).

The functional groups of samples were determined using Fourier transform infrared (FTIR) spectroscopy (Spectrum GX, and Perkin- Elmer) with 20 scans and a resolution of 4 in the spectral range of 400–4,000 (Zhang et al. 2019). Scanning electron microscopy (SEM) (Leo 1455 VP model, made in Germany) was applied to determine the surface morphology of the Mg-HCSB.

The pH _(pzc) value of Mg-HCSB was determined using 100 mL Erlenmeyer flasks containing 50 mL of 0.01 M of sodium chloride, to which 0.1 g of Mg-HCSB was added. The initial pH (pH_i) values were adjusted within the range of 2–11, by adding 0.1 M of hydrochloric acid or 0.1 M of sodium hydroxide. The flasks were agitated at an adjusted temperature of 22 °C ± 2 °C and an agitation speed of 200 rpm. After 24 h, the mixtures were filtered and the final pH values were gauged in the liquid phase. The point of zero charge (pH _(pzc)) was estimated from a plotted graph with ΔpH (pH_i − pH_f) as the vertical axis and the initial pH values (pH_i) as the horizontal axis. In the graph, the point of zero charge of the Mg-HCSB is which ΔpH = 0.

The ANN is a data-driven model of the artificial intelligence family that has attracted tremendous attention due to its capabilities in various fields. This model has been used to predict of the wastewater treatment processes in the last decade (Netto et al. 2021; Maurya et al. 2022).

The GEP method combines algorithms and genetic programming where genes or chromosomes are encoded as linear strands of a constant length. In solving problems using GEP, the algorithm determines the main model variables, such as the gene length and population size. The primary chromosomes are identified randomly or by considering the input information of the problem. Chromosomes are evaluated using fitting functions. Evolution stops if the desired solution is reached or a certain number of generations are reached and the best solution is provided. If the stopping conditions are not found, elitism is selected, and the rest of the solutions are left to the selection process. This process is repeated for several generations and the population quality also relatively improves as the generation progresses (Hemmat-Sarapardeh et al. 2020). In the present study, GenXproTools 4.0 Advanced Edition software was used for the GEP model, and the main arithmetic function (addition-subtraction-multiplication-division-trigonometry and other mathematical functions) were applied to predict the adsorption capacity. The maximum number of generations for model training in this software was between 5,000 and 10,000, and the other parameters used in the GEP model are presented in Table 1. These parameters have been selected through trial and error to obtain the least amount of error is obtained between the actual results and the forecast. The schematic of the process of the GEP model is presented in Figure 1(b).

A support vector machine is one of the most common forecasting techniques in data mining. This model makes its predictions using linear and non-linear combinations on a set of training data called support vectors. The construction process consists of two stages of training and testing, which in the training stage learns the model and the relationships between the data. It can then estimate new data from the detected pattern. Unlike classical statistical methods, this method does not require special assumptions about the data. A very strong assumption in the support vector machine is that the data is linearly separable. Therefore, a dividing line is drawn using special algorithms. Then the two parallel borderlines become the official dividing line. These lines are equidistant from the dividing line enough to collide with the data. The data that collide with this line are called support vectors. For the production of the SVR model in this study, the software package STATISTICA was used. According to the shape of the error function, the SVM models are divided into four types, and in this study, Regression SVM Type 1, which is also known as epsilon-SVM regression, was used. Also, Radial Basis Function kernels were chosen for the SVM model because it has a finite response in the entire X-axis range. Also, it has worked well in times when there was no knowledge of data prediction. to obtain the final model of SVM, the combination of stop at error (accuracy) options and the number of iterations or accuracy (whichever is reached first) was used. Other parameters used in the SVM model are shown in Table 2. Also, the schematic of the SVM model showed in Figure 1(c).

The dataset is considered an important feature in determining the performance of soft computing techniques. Typically, large datasets lead to better classification performance, and small datasets may cause overfitting (Althnian et al. 2021). No one can say exactly how much data is needed for predictive modeling, and this is an intractable problem whose answer is to be discovered through empirical research. For example, the amount of data for machine learning depends on many factors such as the complexity of the problem and the complexity of the learning algorithm. One of the cited sources to determine the required amount of data is to refer to the results of similar studies. Studies tell us how much data is needed to use a particular technique. In many cases, the results of several studies can be averaged. In the research work by (Ullah et al. 2020a), a data set with 48 experimental data was used for modeling purposes by ANN to estimate the Pb (II) adsorption on rice husks. Ullah et al. (2020b), used the data set with 24 data to predict the adsorption of Zn (II) on rice husks digested with nitric acid by ANN. In this study, the characteristics of the experimental data used in the artificial intelligence models are shown in Table 3.

The elemental compositions of the samples are summarized in Table 4. The carbon content increased from 44.4% (in the SB) to 49.54% (in the HCSB) and 47.48% (in the Mg-HCSB). In the hydrochars, the H and O concentrations and atomic ratios of O/C and H/C are lower than those values detected in SB. These results might be attributed to the complex reaction network (dehydration, decarboxylation, and dehydrogenation) of the SB during hydrothermal carbonization (Qian et al. 2018; Zhang et al. 2020).

Figure 4(d) presents the changes in nitrate decontamination percentages due to varying the adsorbent Mg-HCSB dosage. As demonstrated in Figure 4(d), increasing the Mg-HCSB dosage from 0.2 to 10 g L⁻¹ resulted in a noticeable increase in nitrate removal efficiency (from 13.18% to 76.51%). The improved removal percentages might be related to increased sites in Mg-HCSB and the total surface area associated with higher dosages (He et al. 2020). Conversely, the adsorption capacity declined with higher Mg-HCSB dosages, achieving the lowest rate at 10 g L⁻¹. Based on these results, an optimum adsorbent dosage was 1 g L⁻¹ (Mg-HCSB), balancing the adsorption capacity and efficiency. As a function of temperature, nitrate adsorption by Mg-HCSB noticeably decreased as temperature increased. For the studied concentration of 100 mg L⁻¹, increasing the temperature from 20 °C to 40 °C decreased the nitrate removal percentage from 56.48% to 43.20%. This outcome indicates exothermic properties in nitrate uptake by Mg-HCSB, in agreement with the results reported in previous literature (Katal et al. 2012; Wu et al. 2016).

The negative correlation between adsorption capacity and the Mg-HCSB dosage could be related to unsaturated sites, which are more available at higher adsorbent doses. Increasing the adsorbent dosage provides more adsorption sites, whereas the nitrate concentration is constant. Therefore, the adsorption proportion for each gram of adsorbent is decreased at a higher Mg-HCSB dosage. In addition, a decrease in adsorption capacity might be due to an overlap of adsorption sites that increase the diffusion path length and consequently decrease the adsorption capacity. The negative correlation of the pH with the adsorption capacity of Mg-HCSB for nitrate is approved based on the results and reasons stated in Section 3.2. The statistical parameters of the models are presented in Table 6. The results indicate that ANN model provides a better fit as demonstrated by a lower RMSE and higher R² value.

In this study, HCSB was modified using MgCl₂ to remove nitrate from water. There are several reasons why sugarcane bagasse was chosen for hydrochar production, why hydrochar is a good adsorbent, why adsorption is an effective method, and why artificial intelligence tools were used to predict it. Sugarcane is one of the most significant agricultural products that grow in tropical countries. Moreover, SB is a by-product of the sugar industry generated after water is extracted from sugarcane. Furthermore, SB has been introduced as a solid agricultural waste product and its management is challenging due to excess production (Akiode et al. 2015; Kumar et al. 2021). Therefore, creating valuable products from sugar industry waste is imperative, from an environmental viewpoint and biorefinery perspective (Kumar et al. 2021).

Hydrochar production is among the cost-effective and environmentally friendly methods recently developed to manage agricultural waste (Wang et al. 2020). Hydrochar is popular because it is quickly and easily produced, cheap, and clean. Furthermore, hydro cardiogram is rich in organic matter. Applied as a biofertilizer, it increases soil organic carbon, stabilizes aggregates, and reduces greenhouse gas emissions (Hou et al. 2020).

The increase in agricultural effluents due to the development of agriculture has led to the introduction of a wide range of chemical pollutants, pesticides, and nitrate fertilizers into water sources. The high concentration of these pollutants in human drinking water causes many problems such as cancer, hormonal disorders, heart diseases, and kidney and liver damage, especially in children and pregnant women. Therefore, research on the effective and cost-effective removal of various pollutants from water sources is of great importance for public health. Another issue that should be considered is that the world is under pressure due to the increase in population and the increase in demand for water and food. It should be noted that our drinking water resources are very limited so only 1% of the total water on earth is fresh and usable. Therefore, it is very important to recycle and purify polluted water and reuse it. There are various methods for purifying polluted water, surface absorption has advantages over other methods due to its simple design and can include low investment in terms of initial cost and required land (Rashed 2013). Adsorption is a process that uses porous solid materials (absorbents) to separate contaminants from contaminated water. The amount of adsorption is often predicted using adsorption isotherms, which are mathematical expressions developed to describe equilibrium relationships between adsorbent and pollutant. Before using the adsorption isotherm, its coefficients must be adjusted using experimental data (Mahmoodi et al. 2018). To provide experimental data so that they can be used in adsorption isotherms is very time-consuming and expensive. Because it is necessary to investigate the effect of a wide range of different parameters on the adsorption process. In addition, the adsorption capacity of an adsorbent depends on the active sites (adsorbent properties) and the intensity of the adsorption process, which is difficult to detect (Wang et al. 2021).

The findings show that artificial intelligence tools have provided advantages over conventional mathematical modeling, such as: requiring less time to develop the model, avoiding extensive experimental work to formulate a non-linear relationship, and the ability to learn complex relationships regardless to understand the structure of the model (Alam et al. 2022). The application of artificial intelligence methods in predicting of the capacity of adsorbents, can reduce the cost, effort, and time of the optimization process. In addition to the experimental and molecular modeling, the application of artificial intelligence techniques can simulate the adsorption process to understand the effect of fundamental parameters on the removal efficiency and adsorption capacity. It should be mentioned that despite the several advantages offered by artificial intelligence tools, there are still shortcomings that need to be addressed in order to realize the potential of artificial intelligence tools in practical water treatment applications (Alam et al. 2022).

In this study, HCSB was modified using MgCl₂ to obtain a new and unique adsorbent. The effect of temperature, contact time, initial nitrate concentration, adsorbent dose, and solution pH on nitrate uptake by Mg-HCSB were investigated. Laboratory data obtained from adsorption experiments were modeled using artificial intelligence methods. The contact time, initial nitrate concentration, temperature, adsorbent dose, and solution pH were introduced as input variables and adsorption capacity and removal efficiency were introduced as target variables. The results demonstrated that the ANN model is more accurate in predicting the adsorption capacity due to the lower RMSE and higher R² value. The best conditions for nitrate adsorption by Mg-HCSB occurred at an initial pH of 2, a temperature of 22 °C ± 2 °C, an adsorbent dosage of 1 g L⁻¹, and after 8 h of contact time. The Langmuir isotherm was selected as the best fitting model. Using the Langmuir model, the amount of nitrate adsorbed by Mg-HCSB was 68.296 mg g⁻¹. A neural network needs a suitable amount of quality training data to produce viable results. Sudden fluctuations in input variables due to changes in operational parameters and water quality (concentration of other pollutants), which leads to the development of low-quality data, is one of the concerns of using artificial intelligence models in this research.

The authors of this paper especially thank the Iran National Science Foundation (INSF) for funding this work, through project No. 97013515 and the Shahid Chamran University of Ahvaz, Iran, for the laboratory facilities.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.