Application of machine learning in predicting the adsorption capacity of organic compounds onto biochar and resin

https://doi.org/10.1016/j.envres.2022.112694Get rights and content

Highlights

  • Kriging-LFER model is proposed for predicting adsorption properties.

  • Sensitivity analysis provides the most sensitive parameter of resin are V and logCe.

  • Uncertainty analysis shows that the range of logKd of resin is from 0.062 to 0.189.

Abstract

Detailed prediction of the adsorption amounts of organic pollutants in water is essential to the clean development and management of water resources. In this study, Kriging and polyparameter linear free energy relationship model are coupled to predict adsorption capacity of organic pollutants by biochar and resin. It's based on 1750 adsorption experimental data sets which contains 73 organic compounds on 50 biochars and 30 polymer resins. The Kriging-LFER model shows better accuracy and predictive performance for adsorption (R2 are 0.940 and 0.976) than the published NN-LFER model (R2 are 0.870 and 0.880). Local sensitivity analysis method is adopted to evaluate the influence of each variable on the adsorption coefficient of resin and find out that top sensitive parameters are V and log Ce, to guide parameter optimization. Data's uncertainty analysis is presented by Monte Carlo method. It predicts that the adsorption coefficient will range from 0.062 to 0.189 under the 95% confidence interval. The Kriging-LFER model provides great significance for understanding the importance of various parameters, reducing the number of experiments, adjusting the direction of experimental improvement, and evaluating the fate of organic pollutants in the environment.

Introduction

The severe shortage of clean water poses huge challenges for our modern social development (Ma et al., 2020). The organic pollutant of water such as personal care products, pesticides and food additives brought a big threaten to water resource and environmental (Rojas and Horcajada, 2020). Regarding the water restoration, physical adsorption is one of the promising strategies to remove organic pollutants on account of cost, simplicity, and energy considerations (Shi et al., 2021). Numerous adsorbents, including carbon nanotubes, granular activated carbons, Graphene nanosheets and biochars have been reported to be able to remove those organic pollutant from water (Bunmahotama et al., 2015; Ersan et al., 2019; Wu et al., 2013). However, It is clear that currently used this technology are not efficient, especially against novel chemicals and adsorbents, and has brought about scarcity in data for adsorption of compounds on adsorbents (Chen et al., 2011a; Kuncek and Sener, 2010; Luo et al., 2019; Zheng et al., 2019). The establishment of efficient prediction models can make full use of existing data and replace some labor-intensive adsorption experiments (Zhang et al., 2020a). Polyparameter linear free energy relationship (pp-LFER) model is a popular model used in predicting adsorption of organic compounds (Qi et al., 2020). The molecular structure parameters of pp-LFER model can explain the distribution of organic pollutants between solid and liquid phases from the perspective of intermolecular force (Zhu et al., 2019a). The size of pp-LFER model coefficients can distinguish the difference in force between the molecules of different systems, and then clarify adsorption mechanism of organic compounds making a correct environmental risk assessment. The pp-LFER model is as follows Eq. (1). The meaning of the symbols of pp-LFER model can be seen in Table S1. However, pp-LFER model tends to ignore the effect of adsorbent properties and leads to large prediction errors. In addition, it usually needs to establish a multilinear regression (MLR) for each equilibrium concentration Ce, and the prediction based on the MLR is limited to the concentration level involved in the modeling (Zhang et al., 2020a). Considering limitations of the above method, the machine learning model combining basic properties of adsorbents and compounds can provide a helpful solution.logKd=eE+sS+aA+bB+vV+c

Machine learning (ML), an interdisciplinary subject involving various fields, has been applied to adsorption research by scholars at present (Kobayashi et al., 2020). The theory of ML is mainly to design and analyze some algorithms that let computer to “learn” automatically. Based on the principles of statistics, ML algorithms can automatically analyze the structure of existing data and mine regulation, to judge and predict for the unknown samples (Mazaheri et al., 2017). In terms of organic pollutants adsorption, ML methods (such as artificial neural networks (ANN), support vector machines (SVM), random forest (RF), etc.) have gradually become the key method of research (Sahu et al., 2019). ML method can not only optimize the adsorption parameters (Rahman et al., 2019; Zhang and Pan, 2014), but also has more obvious effect than the traditional regression methods in simulating single-component/multi-component adsorption (Panapitiya et al., 2018; Zhang et al., 2019b). ANN model was used to study the ability of biochar to predict the adsorption of pollutants. It is proved that the overall RMSE and R2 values obtained by ANN algorithm were acceptable, but some of the prediction performance was poor (Yang et al., 2014). Least square support vector machine (LS-SVM) and ANN was proved that can promote the adsorption efficiency of methylene blue fuel (Asfaram et al., 2016). Widely used feedforward backpropagation ANN and RF were compared in performance, and their applicability was evaluated for 336 energy consumption predictions. It was found that the two models perform well in training and validating data (Ahmad et al., 2017). Previous studies found that incorporating the surface area of describing the adsorbent properties into a deep neural network model to predict the fitted parameters of the Freundlich isotherm largely solved the problem of the MLR limiting equilibrium concentration (Sigmund et al., 2020). In addition, ANN-LFER method based on this strategy provides a solution to solve the limitations of existing single solute adsorption prediction model, but the prediction accuracy still needs to be further improved (Zhang et al., 2020a).

Compared to the above methods, Kriging is an accurate interpolation method, which has received extensive attention in modeling in recent years (Xiao et al., 2020). Originated from geostatistics, Kriging model has been applied to various optimization problems and structural reliability analyses (Huang et al., 2016). Kriging model has the characteristics of high prediction accuracy, less time-consuming and strong robustness (Zhao et al., 2016). At present, it has been widely used in meteorology, geography, and computational science, but there are few reports on the organic compounds’ adsorption (Zhao et al., 2021).

To address the limitations of the existing models, Kriging-LFER model is established based upon the collected data. Moreover, the sensitivity analysis is used to evaluate the influence of each variable on the output of the model. Finally, the range of adsorption coefficient is effectively determined by uncertainty analysis, and the most likely results of the model are predicted. Kriging-LFER model established in this study can not only reduce the experimental workload, but also rapidly predict adsorption efficiency of biochar and resin for organic compounds according to the basic properties of adsorbent and compounds, which is of great significance for understanding the importance of various parameters, adjusting, and improving the direction of experiment, and environmental governance.

Section snippets

Data collection

In this study, biochar and polymer resin are selected as the target adsorbent, meanwhile, high-quality, representative raw data are selected to establish a prediction model. To compare with developed models in previous studies, experimental data are selected the same as previous studies (Zhang et al., 2020a). Supporting Information about the collected data is shown in Table S2.1-S2.2. 1750 adsorption data points related to adsorption isotherms are excavated. These concentration points, which

Statistical results of biochar and resins characteristics and pearson correlation matrix analysis

A description of the statistical distribution of variables for 50 biochar and 30 polymeric resin is presented in Fig. 1. Box plot uses the maximum value, upper and lower quartile, median, and minimum value to describe data from top to bottom. The data set is based on quartiles and interquartile ranges to measure the variability, while points outside the inner limit are drawn with diamonds. The results of boxplot show that the value of the BET area is 2–3 orders of magnitude higher than that of

Conclusion

Machine learning model can effectively predict the adsorption efficiency of biochar and resin for organic pollutants in aqueous solution, and it can achieve lower prediction errors (R2 = 0.940, RMSE = 0.037 and R2 = 0.976, RMSE = 0.019). Compared with ANN-LFER, the prediction accuracy of Kriging model is improved by 7% and 9.6%, respectively. In addition, sensitivity analysis provides the priority of parameter adjustment (V>logCe>Vt>BET>B>S>E>A), having important reference value for optimizing

Declaration of competing interest

The authors declared that they have no conflicts of interest to this work.

We declare that we do not have any commercial or associative interest that represents a conflict of interest in connection with the work submitted.

Acknowledgements

The authors appreciate the financial support by National Natural Science Foundation of China (No. 41807196), Postdoctoral Science Foundation of China (2021T140104, 2018M641793), Excellent Youth of Heilongjiang Province Natural Science Foundation (YQ2021D003), Postdoctoral Science Foundation of Heilongjiang Province of China (No.LBH-Z19002).

References (48)

  • P.S. Pauletto et al.

    Single and competitive dye adsorption onto chitosan-based hybrid hydrogels using artificial neural network modeling

    J. Colloid Interface Sci.

    (2020)
  • X. Song et al.

    Variance-based sensitivity analysis of a forest growth model

    Ecol. Model.

    (2012)
  • Z.L. Sun et al.

    LIF: a new Kriging based learning function and its application to structural reliability analysis

    Reliab. Eng. Syst. Saf.

    (2017)
  • M. Wu et al.

    The sorption of organic contaminants on biochars derived from sediments with high organic carbon content

    Chemosphere

    (2013)
  • X.Y. Zhang et al.

    Achieving no spent brine discharge in an anion exchange resin-based fixed-bed process for typical DOM removal

    Chem. Eng. Res. Des.

    (2019)
  • Y. Zhang et al.

    Maximizing natural frequencies of inhomogeneous cellular structures by Kriging-assisted multiscale topology optimization

    Comput. Struct.

    (2020)
  • Y.Y. Zhang et al.

    Modeling batch and column phosphate removal by hydrated ferric oxide-based nanocomposite using response surface methodology and artificial neural network

    Chem. Eng. J.

    (2014)
  • Y. Zhao et al.

    A Kriging surrogate model coupled in simulation-optimization approach for identifying release history of groundwater sources

    J. Contam. Hydrol.

    (2016)
  • Y.F. Zhao et al.

    Prediction of adsorption properties for ionic and neutral pharmaceuticals and pharmaceutical intermediates on activated charcoal from aqueous solution via LFER model

    Chem. Eng. J.

    (2019)
  • M.Q. Zheng et al.

    Adsorption desulfurization performance and adsorption-diffusion study of B2O3 modified Ag-CeOx/TiO2-SiO2

    J. Hazard Mater.

    (2019)
  • E. Anane et al.

    Output uncertainty of dynamic growth models: effect of uncertain parameter estimates on model reliability

    Biochem. Eng. J.

    (2019)
  • A. Asfaram et al.

    Statistical experimental design, least square-support vector machine (LS-SVM) and artificial neural network (ANN) methods for modeling of facilitated adsorption of methylene blue dye

    RSC Adv.

    (2016)
  • S.S. Bai et al.

    Self-assembly biochar colloids mycelial pellet for heavy metal removal from aqueous solution

    Chemosphere

    (2020)
  • Y. Kobayashi et al.

    Prediction of soil adsorption coefficient in pesticides using physicochemical properties and molecular descriptors by machine learning models

    Environ. Toxicol. Chem.

    (2020)
  • Cited by (24)

    View all citing articles on Scopus
    1

    Ying Zhao and Da Fan contribute equally to this work and should be considered co-first authors.

    View full text