Researchers from MIPT and Sichuan University have developed a method called SWNN to reconstruct 3D models of porous materials from a single 2D image. The approach can generate models containing hundreds of millions of cells, as detailed in the journal ScienceDirect.
The method utilizes neural networks combined with the Wasserstein distance metric. Unlike traditional deep learning models that rely on large datasets, SWNN works with a single sample. It allows for precise control over properties like porosity and the creation of spatially heterogeneous structures based on specific masks.
Experiments demonstrated that the reconstructed models are statistically and physically indistinguishable from real samples, specifically regarding thermal conductivity and fluid permeability. The approach outperformed existing generative models in quality while requiring fewer computational resources.
This technology facilitates the design of materials with targeted characteristics, such as porous anodes and cathodes for batteries, catalysts, and filters. The research team is currently exploring additional AI-based and non-AI methods for material structure analysis and generation.
