Neural network-optimized imaging for classifying lignin-based polyurethane foams: Linking molecular composition to cellular microstructure using advanced machine learning

Abstract

In recent years, there has been growth in machine learning (ML) applications in polymer science. However, applying ML strategies to solve problems faced by polymer chemists remains in its infancy. A critical challenge is designing polyurethane (PU) foams with tailored microstructures and properties, a process still largely reliant on time consuming trial-and-error. This study introduces Convolutional Neural Networks (CNNs), an optimized ML algorithm for image processing, to explore structure-property relationships in biobased PU foams derived from lignin hydrogenolysis oil. The dataset included specimens with varying compositions, characterized by compression testing and scanning electron microscopy images taken before and after compression. A 30:70 training-to-validation split was used for model development. The CNN optimized model for classification achieved excellent performance, to identify PU foam composition based on geometric features. For validation, the CNN optimized model was compared against the "ResNet-50″ model. Across both compressed and uncompressed datasets, the presented CNN optimized model consistently outperformed "ResNet-50″, achieving higher accuracy (up to 0.99) and high precision (0.97). The findings demonstrate the method's reliability, especially with data from compressed foams. This study underscores the transformative potential of ML in accelerating material design, offering a streamlined approach for developing PU foams with customized microstructures and enhanced performance.