Deep spatial representation learning of polyamide nanofiltration membranes

Abstract Machine learning overfitting caused by data scarcity greatly limits the application of chemical artificial intelligence in membrane materials. As the original data for thin film polyamide nanofiltration membranes is limited, here we propose to extract the natural features of monomer molecular structures and rationally distort them to augment the data availability. This few-shot learning method allows a chemical engineering project to leverage the powerful fit of deep learning without big data at the outset, which is advantageous over traditional machine learning models. The rejection and flux predictions of polyamide nanofiltration membranes are practiced by the molecular augmentation in deep learning. Convergence of loss function indicates that the model is effectively optimized. Correlation coefficients over 0.80 and the mean relative error below 5% prove an accurate prediction of nanofiltration performance. The success of predicting nanofiltration membrane performances is widely instructive for molecule and material science.