Gradient boosting decision tree algorithms for accelerating nanofiltration membrane design and discovery

Abstract

<p>Interfacial polymerization is the most widely used strategy for nanofiltration membrane fabrication. Despite extensive research on this technology, further improvement in permeance and salt rejection is still essential due to its multidimensional characteristics, including the types of membrane material and the conditions of membrane optimizing fabrication. Herein, we applied four gradient boosting decision tree algorithms to precisely identify the candidate monomers (represented by the molecular descriptors) and their fabrication conditions. The result of the model evaluation indicated the Extreme Gradient Boosting (XGBoost) algorithm had the best predictive performance in accuracy and generalization in predicting membrane permeance and salt rejection, with the corresponding determination coefficients on the test set being 0.76 and 0.88. Shapley additive explanation analysis showed that the aqueous monomer concentration was the most influential fabrication condition in membrane performance. Besides, the partition coefficient (Log P) and topological polar surface area were the most important molecular descriptors in water permeance and salt rejection, respectively. Overall, this study proposed innovative machine learning algorithms to disentangle the multidimensional interactions of various influencing factors on membrane performance, thus initiating a paradigm shift in the development of high-performance nanofiltration membranes.</p>