Using Machine Learning to Correlate Microstructure with Diffusion, Flow and Wetting: New Results Regarding Connectivity and Bottleneck Effects
Please login to view abstract download link
Using TEM, FIB-SEM imaging, X-ray tomography, and small-angle scattering, we analyze the microstructure of porous, soft materials across multiple length scales. With data from the packaging, pharmaceutical, and hygiene industries, we build virtual material models to explore links between microstructure and properties of diffusion, flow, and wetting. Through extensive simulations on diverse structures—such as fiber networks, polydisperse packings, and polymer-based materials—we apply machine learning to identify correlations between microstructure and these properties. This extends previous work by including more diverse structures and extending the examined relationships to wetting [1]. Our findings reveal that simple microstructural features, such as connectivity and bottleneck effects, explain much of the variability for both ordered and disordered, homogeneous and inhomogeneous structures. Microstructural descriptors capturing these effects, along with 3D visualization tools, are made available in the open-source software Mist [2]. We demonstrate the method on real materials with breathable films used in hygiene products, highlighting how poorly connected pore networks affect diffusion, and with paper coatings, showing how the coating structure impact flow and wetting.