Persistence is All You Need -- A Topological Lens on Microstructural Characterization

TL;DR

This paper introduces a novel method combining computational topology and deep learning to rapidly and accurately extract key microstructural descriptors from 3D materials, aiding material design.

Contribution

It presents a unified workflow that uses persistence images and neural networks to predict multiple microstructural descriptors from synthetic data.

Findings

Achieved an average R^2 of 0.84 in predictions

Demonstrated high correlation with Pearson r ~ 0.92

Provided a scalable tool for microstructural characterization

Abstract

The microstructure critically governs the properties of materials used in energy and chemical engineering technologies, from catalysts and filters to thermal insulators and sensors. Therefore, accurate design is based on quantitative descriptors of microstructural features. Here we show that eight key descriptors can be extracted by a single workflow that fuses computational topology with assembly-learning-based regression. First, 1312 synthetic three-dimensional microstructures were generated and evaluated using established algorithms, and a labeled data set of ground-truth parameters was built. Converting every structure into a persistence image allowed us to train a deep neural network that predicts the eight descriptors. In an independent test set, the model achieved on average R^2 ~ 0.84 and Pearson r ~ 0.92, demonstrating both precision and generality. The approach provides a unified and scalable tool for rapid characterization of functional porous materials.