This study presents a novel approach for predicting the characteristics of porous materials, such as permeability and filtration efficiency, using 2D images. The project is driven by requests from the Center for Filtration Research (CFR) members, addressing the challenge of analyzing filter structures more efficiently. Traditional methods like computational fluid dynamics (CFD) simulations are resource-intensive, whereas 2D image analysis offers a quicker, more storage-efficient alternative. The research leverages convolutional neural networks (CNNs) to interpret 2D images and predict properties of these complex structures.
A significant portion of the research focuses on building a large dataset using the GeoDict software, which simulates 3D porous structures and generates corresponding characteristics for machine learning training. The dataset includes various geometries such as spheres, fibers, and polyhedrons, ensuring the model's generalizability.
A key challenge addressed by the project is how to accurately represent 3D structures using 2D images, given the loss of depth information. To mitigate this, the study explores different methods for converting 3D structures into 2D images. The CNN model has been trained to identify these patterns and predict material properties based on 2D representations. Results show that while the model performs well with isotropic structures, its generalizability can be improved when anisotropic structures are included in the training dataset. Further refinement is planned, with ongoing work to validate the model using real-world SEM images and to enhance its ability to interpret more complex structures.
The potential applications of this research offer meaningful advantages for industries that depend on filtration systems, such as environmental engineering, healthcare, and manufacturing. By leveraging 2D images to predict properties like permeability and filtration efficiency, this approach aims to provide a more efficient method for evaluating filter materials. This could help streamline product development by reducing reliance on more complex simulations and offering a faster way to assess material performance. Once validated with real-world SEM images, this method may enhance material selection processes, supporting companies in refining their designs and improving filter performance. While still in development, this research presents a promising step toward more efficient and cost-effective filtration solutions.