Deep neural networks for modeling fouling growth and flux decline during NF/RO membrane filtration

https://doi.org/10.1016/j.memsci.2019.06.004Get rights and content

Highlights

  • A deep learning-based fouling model was developed using a deep neural network.

  • Membrane fouling based on flux decline and fouling layer thickness were evaluated.

  • The model had the highest performance amongst the tested models.

  • This study provided a new tool for simulating fouling using optical image data.

Abstract

Mathematical models have been developed to obtain a better understanding of membrane fouling mechanisms. However, those models could not simulate the membrane fouling behaviors accurately because of the large number of fitting parameters related to feed water quality and flow pattern in a membrane filtration system. In this study, we developed a deep neural network (DNN) to model membrane fouling during nanofiltration (NF) and reverse osmosis (RO) filtration using in-situ fouling image data from optical coherence tomography (OCT). The performance of the DNN model was compared with that of existing mathematical models. In total, 13,708 high-resolution fouling layer images were used to develop the DNN model and validate the model performance. The DNN model was trained to simulate both organic fouling growth and flux decline, and it reproduced two- or three-dimensional images of the organic fouling growth. The DNN model demonstrated better predictive performance than the existing mathematical models. It achieved an R2 value of 0.99 and RMSE of 2.82 μm for the fouling growth simulation and R2 of 0.99 and RMSE of 0.30 Lm−2h−1 for the flux decline simulation. Therefore, the data-driven approach is an alternative way to model the membrane fouling and flux decline processes under high-pressure filtrations.

Introduction

Membrane filtration is widely used for drinking water and wastewater treatment processes [1]. Four different types of membrane filtration techniques such as microfiltration, ultrafiltration, nanofiltration, and reverse osmosis are applied to remove salts or organic matter depending on the need [2]. However, fouling is regarded as one of the critical issues in the membrane filtration process because it accelerates the deterioration of water quality, thereby increasing the water production cost [3].

Numerous studies have been conducted to understand and control fouling on membrane surfaces by considering various parameters such as channel flow velocity, organic concentration, and ion strength [4,5]. Besides experimental studies, membrane models (i.e., mathematical or data-driven models) have been developed to obtain a better understanding of membrane fouling and flux decline so that the optimum maintenance conditions can be found [6]. Mathematical models, however, were not very successful in accurately simulating the membrane fouling behaviors due to its complexities (i.e., heterogeneity of feed water quality, unexpected bacterial growth/death, and complex flow pattern, etc.) [7].

Therefore, a data-driven model would be an alternative way to overcome these complexities and achieve higher predictive accuracy for membrane fouling and flux decline [8]. This is because the model is optimally trained to achieve a good-fit on the actual monitoring data using nonlinear and kernel functions [9]. In particular, artificial neural networks (ANNs) have been used as a modeling tool for membrane filtration systems [10]. One of the advantages of applying ANNs is that it does not require the solving of complex partial differential equations, and requires only available explanatory and target variables in a membrane process [11].

Deep neural networks (DNNs), which are considered as an advanced version of ANN, have been proposed as a promising state-of-the-art technique in many different fields, including speech recognition, image recognition, remote sensing, and climate forecasting [[12], [13], [14]]. The most distinguishable feature of DNN from ANN is that it automatically extracts representations from the given big data using multiple inter-connected layers [15]. Due to this feature, current environmental studies utilize DNN in order to optimize the performance of treatment plants [9,16]; however, the membrane science field still lacks the knowledge to apply DNN in studies such as flux decline and fouling behaviors.

Currently, optical coherence tomography (OCT), a non-destructive and high-speed scanning technology, is used as a fouling monitoring device [17,18]. OCT provides extremely fine-scale data on the membrane surface (<7 μm) and enables real-time direct observation of fouling on a membrane by providing numerous images of the membrane surface within a short time (120 frames/s). However, OCT methodology appears to be difficult to be utilized for continuous monitoring in the field conditions [19]. In addition, the monitoring ability has a limitation in improving operating performance since it can represent only current state of fouling, but does not provide a predictive information for the future. Predicting membrane fouling should be very useful for optimizing the operation of water treatment processes.

With this background, this study attempts to apply DNN to model membrane fouling for the first time. This study aims 1) to develop a DNN model for predicting organic fouling growth and flux decline of NF/RO membranes using high-resolution fouling layer images, and 2) to evaluate the performance of the model with OCT images. In addition, we produced a 3-dimensional video clip on organic fouling growth (available online). We hope this study will provide further insights into the membrane science with a DNN approach.

Section snippets

Experimental setup and data acquisition

The experimental data was taken from our previous study [19]. In this study, membrane fouling was performed using an NF membrane (NE 90, Toray, Japan) and an RO membrane (RE SHF, Toray) with 10 mg C/L of humic acid (Sigma-Aldrich, USA) and 10 mM of calcium ions (Sigma-Aldrich, USA). The effect of calcium ions on fouling layer growth was monitored by using a surface scanning device OCT (SD-OCT, Oz-tec, Korea). The feed water temperature was maintained around 20 °C. The scan area was a region of

The effect of deep learning on flux decline estimation

Fig. 3 compares the observed and simulated flux variations of each OCT image, depending on the membrane types. Model calibration was performed to estimate the model accuracy. In the calibration set, the R2 and RMSE values were 0.98 and 1.44 Lm−2h−1, respectively. The model was validated by comparing the simulated and observed data, except for the images used for training the model. In the validation set, the R2 and RMSE values were 0.98 and 1.52 Lm−2h−1, respectively. In the test set, the R2

Conclusions

The fouling model presented and discussed in this study provides a more precise and remarkable performance for estimating both flux decline and fouling layer growth, simultaneously. The conclusions we obtained are summarized as follows:

  • Deep learning was able to utilize substantial amount of image data using convolutional neural network for developing fouling model that estimates flux decline and fouling layer growth.

  • The fouling model successfully estimated the flux decline of two different

Acknowledgements

This work was supported by Korea Environment Industry & Technology Institute (KEITI) through Industrial Facilities & Infrastructure Research Program, funded by Korea Ministry of Environment (MOE) (RE201901123), and the National Research Foundation of Korea Grant funded by the Korean Government (MSIP) (No. NRF2015R1A5A7037825).

References (49)

  • F. Harrou et al.

    Statistical monitoring of a wastewater treatment plant: a case study

    J. Environ. Manag.

    (2018)
  • S. Park et al.

    Investigating the influence of organic matter composition on biofilm volumes in reverse osmosis using optical coherence tomography

    Desalination

    (2017)
  • J. Park et al.

    Evaluation of fouling in nanofiltration for desalination using a resistance-in-series model and optical coherence tomography

    Sci. Total Environ.

    (2018)
  • J. Park et al.

    Modeling of NF/RO membrane fouling and flux decline using real-time observations

    J. Membr. Sci.

    (2019)
  • S.-C.B. Lo et al.

    Artificial convolution neural network for medical image pattern recognition

    Neural Network.

    (1995)
  • F. Faridirad et al.

    Modeling of suspension fouling in nanofiltration

    Desalination

    (2014)
  • C.-C. Ho et al.

    A combined pore blockage and cake filtration model for protein fouling during microfiltration

    J. Colloid Interface Sci.

    (2000)
  • J. Wu et al.

    Modeling of the submerged membrane bioreactor fouling by the combined pore constriction, pore blockage and cake formation mechanisms

    Desalination

    (2011)
  • G. Bolton et al.

    Combined models of membrane fouling: development and application to microfiltration and ultrafiltration of biological fluids

    J. Membr. Sci.

    (2006)
  • L. Huang et al.

    Fouling of membranes during microfiltration of surimi wash water: roles of pore blocking and surface cake formation

    J. Membr. Sci.

    (1998)
  • A. Seidel et al.

    Coupling between chemical and physical interactions in natural organic matter (NOM) fouling of nanofiltration membranes: implications for fouling control

    J. Membr. Sci.

    (2002)
  • C.Y. Tang et al.

    Colloidal interactions and fouling of NF and RO membranes: a review

    Adv. Colloid Interface Sci.

    (2011)
  • A. Lim et al.

    Membrane fouling and cleaning in microfiltration of activated sludge wastewater

    J. Membr. Sci.

    (2003)
  • K. Katsoufidou et al.

    A study of ultrafiltration membrane fouling by humic acids and flux recovery by backwashing: experiments and modeling

    J. Membr. Sci.

    (2005)
  • Cited by (53)

    View all citing articles on Scopus
    1

    Sanghun Park and Sang-Soo Baek are Co-first authors.

    View full text