Harnessing the power of machine learning methods for the investigation of direct contact membrane distillation systems

Date of Award


Document Type



Edith Cowan University

Degree Name

Doctor of Philosophy


School of Engineering

First Supervisor

Mehdi Khiadani

Second Supervisor

Masoumeh Zargar

Third Supervisor

Abdellah Shafieian


Desalination processes have the potential to significantly address the global water scarcity crisis. This is primarily because approximately 97% of water sources are saline or brackish, underscoring the importance of these processes. Solar-driven membrane distillation (MD) has garnered considerable attention in recent years among various solar-powered water treatment/desalination methods. This attention is due to its distinct advantages, such as operating at low temperatures, having a simple and compact design, the ability to treat highly saline water, requiring minimal membrane mechanical properties, and producing high-purity water. However, solar-driven MD systems face several significant challenges, including low gained-output-ratio (GOR), limited freshwater production, and relatively high water costs. Addressing these challenges requires a precise analysis of individual MD modules and solardriven MD systems. Among MD configurations, the direct contact membrane distillation module (DCMD) features as the most suitable for integration with solar energy. As the core component of solar-driven DCMD systems, analysing the performance of DCMD modules under various operational conditions can provide valuable insights to enhance system performance. Moreover, conducting accurate performance analyses of solar-driven DCMD systems and proposing strategies to improve their efficiency from both an energetic and economic perspective are essential tasks that must be undertaken. To achieve these objectives, this study leverages both experimental and machine learning methods.

In terms of analysing the performance of DCMD modules, novel performance modelling tools were introduced to accurately diagnose their behaviour under various operational conditions. Machine learning models, including artificial neural network (ANN), support vector regression (SVR), and Random Forest (RF), were utilized to predict permeate flux with high accuracy. Comparative analysis demonstrated the superiority of machine learning models over conventional mechanistic models, particularly in assessing the impact of operational parameters such as feed flow temperature and salinity. The results demonstrated that the mean absolute percentage error of the test dataset for ANN, SVR, and the mechanistic modelling approach reached 3.46%, 4.78%, and 7.31%, respectively. Additionally, the machine learning models exhibited significantly lower computational times compared to the stepwise mechanistic modelling approach, emphasizing their computational superiority, especially for large-scale membrane distillation modules.

An in-depth examination was also conducted on DCMD modules subject to organic fouling conditions using both experimental and machine learning methods. The performance of the DCMD module, experiencing organic fouling, was experimentally investigated across a wide range of operational parameters to accommodate the required data for the modelling purpose. Subsequently, ensemble machine learning models, including Random Forest (RF), Extra Trees (ET), and Extreme Gradient Boosting (XGBoost), accurately modelled DCMD module performance under fouling conditions. Results indicated that elevated feed temperatures increased permeate flux, while lower temperatures reduced fouling effects and increased gained-output-ratio (GOR). Furthermore, the study identified feed mass flow rate as the parameter with the most significant impact on both permeate flux and GOR, underscoring the importance of parameter optimization in mitigating organic fouling.

Regarding solar-driven DCMD systems, a novel methodology leveraging machine learning approaches was introduced to assess their dynamic performance. Utilizing time-series neural network models, including multilayer perceptron (MLP) and long short-term memory (LSTM), system’s performance under varying operational and weather conditions was comprehensively analysed. Results highlighted the superiority of the MLP model in dynamically modelling system efficiency, with R 2 test and MAPEtest for the MLP model at 0.99 and 5.453%, respectively, compared to 0.93 and 12.96% for the LSTM model. Additionally, analysis identified specific parameters, such as feed stream temperature and solar intensity, as having the most significant influence on system performance.

Finally, an improved solar-driven DCMD system was proposed and modelled to enhance system performance while minimizing complexity. Through an integrated approach combining experimental data and machine learning models, significant improvements in system efficiency and cost-effectiveness were achieved. Machine learning techniques, including Artificial Neural Networks (ANN) and Support Vector Regression (SVR), accurately modelled system performance, identifying key parameters. Results indicated significant improvements in freshwater productivity and GOR, ranging from 35.39% to 37% and 30.64% to 31.57%, respectively. Moreover, the incorporation of an air-cooled condenser and oil-filled heat pipe evacuated tube collectors (HP-ETC) resulted in a substantial increase in daily efficiency (3538%) and a reduction in cost per liter (CPL) by approximately 20%.



Access Note

Access to this thesis is embargoed until 5th December 2024

Access to this thesis is restricted. Please see the Access Note below for access details.