Speaker
Description
Inverse microstructure design is a persistent challenge in materials engineering because structure-property relations are high-dimensional, stochastic, and expensive to evaluate. As a result, conventional optimization and surrogate-driven workflows often become impractical when the design space is large, and microstructures must satisfy multiple constraints. Here we present PoreFlow, a data-driven framework for high-throughput generation of porous microstructures using continuous normalizing flows (CNFs). PoreFlow conditions the generative process on target properties through a latent representation, enabling efficient sampling of microstructures that meet specified objectives while retaining a continuous, invertible mapping between latent variables and generated structures.
We validate PoreFlow on 3D porous media generation. The framework achieves coefficients of determination above 0.915 for reconstruction and above 0.92 when generating previously unseen samples that satisfy the prescribed targets. In contrast to GAN-based approaches that can suffer from training instability and mode collapse, the flow-based formulation provides stable likelihood-based training and supports more transparent analysis of the latent space. The architecture is modular, allowing the autoencoder component to be replaced to accommodate alternative microstructure parameterizations beyond voxelized images.
PoreFlow provides a scalable pathway for inverse design of porous materials with applications in energy storage, catalysis, and related transport-dominated systems, enabling faster and more reliable exploration of structure space under property constraints.
| Country | United States |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
