Modeling and optimization of metal-organic frameworks membranes for reverse osmosis with artificial neural networks

• Published in: Desalination 2022-06, Vol.532, p.115729, Article 115729
• Main Authors: Yao, Lei, Li, Yong, Cheng, Qisong, Chen, Zhe, Song, Jinling
• Format: Article
• Language: English
• Subjects: Artificial neural network; Mean impact value; Membrane; Metal-organic frameworks; Reverse osmosis
• ISSN: 0011-9164; 1873-4464
• DOI: 10.1016/j.desal.2022.115729

Metal-organic frameworks (MOFs) have recently attracted tremendous attention as membrane materials for desalination owing to their diversified structures, permselectivity, and tunable functionalities. However, the structure-performance relationship of MOF membranes still has not been elucidated. Herein, artificial neural networks (ANN) were developed to form prediction model of MOF thin film nanocomposite (TFN) membrane performances in reverse osmosis applications. Key parameters including MOF size, MOF pore diameter, MOF loading, selective layer thickness, salt concentration, and pressure were collected from literature to predict the water permeability and salt rejection. 5-fold cross-validation and hyperparameter tuning method were employed to acquire a better performing model. When the node structure was 6–9–9–8–2 with the learning rate of 0.001, the developed ANN model attained a remarkable prediction R2 of 90.62%. By introducing mean impact value algorithm into the model, the feature importance of each factor on the membrane performance was also calculated. Rather than the MOF loading and size, proper control of selective layer thickness and MOF pore diameter was key to breaking the permeability-selectivity tradeoff. Finally, ANN proved its ability to predict the membrane performance for water purification and further provide guidance for MOF-based membrane design, or even other TFN membrane design. [Display omitted] •Artificial neural networks are used to optimize and predict MOF TFN membranes.•5-fold cross-validation and hyperparameter tuning were employed.•R2 of 98.08% and 90.62% were achieved for training and prediction.•MIV algorithm was used for variable screening of the BP neural network.•PA thickness, MOF pore diameter, and MOF sizes were the key factors.