COMPUTING - Connecting Membrane Pores and Production Parameters via Machine Learning

This study addresses the pressing issue of producing polymer membranes for drinking water purification in light of climate change. Despite years of research, the precise relationship between pore properties and fabrication parameters remains ambiguous. To address this challenge, the study employs machine learning methods to capture complex relationships between variables of quality. The study includes an in-house dataset of 1970 PS-P4VP-copolymer electron microscopy images with pore structures, along with their fabrication parameters.

A deep learning model is utilized to generate semantic segmentation masks for extracting morphological membrane parameters from microscopy images. Using these generated segmentations allows the extraction of pore-specific parameters, such as size, as well as general structural characteristics, such as pore coverage and distribution, for all images. 

A human-in-the-loop active learning approach is used for training the deep learning model, where human expertise is utilized to optimize the model's prediction by iteratively selecting and labeling the most informative data points for the model to learn from. In this way, the labeling is done efficiently and accurately, resulting in a comprehensive database containing the membrane properties.

Random forests are employed in regression modeling due to their accuracy and robustness in handling complex and high-dimensional datasets. The models can handle both categorical and continuous input variables and are relatively insensitive to outliers and missing data, making them suitable for such real-world datasets. Additionally, the models provide feature importance measures, allowing the identification of the most influential factors in membrane quality.

In parallel to random forests, we used various regression and neural network models to identify the critical parameters that determine the quality of membranes. Our findings from a pipeline of image and tabular data analysis demonstrate precise quality parameter predictions that are in line with experimental findings. The results demonstrate the effectiveness of machine learning in capturing complex relationships between variables for optimizing membrane quality.

Overall, this study showcases a comprehensive approach to improving membrane production for drinking water purification, integrating data collection, deep learning, and regression modeling. The approach provides a promising avenue for future research in this area.