Exploring descriptors for titanium microstructure via digital fingerprints from variational autoencoders

TL;DR

This paper investigates the use of variational autoencoders to create interpretable, low-dimensional microstructural fingerprints from titanium alloy micrographs, enabling better data-driven materials analysis and understanding.

Contribution

It introduces a VAE-based approach for generating continuous, interpretable microstructural descriptors from optical micrographs of titanium alloys, facilitating quantitative analysis.

Findings

VAE fingerprints are smooth and interpolable.

Key microstructural properties correlate with latent space positions.

The approach supports process-structure-property exploration.

Abstract

Microstructure is key to controlling and understanding the properties of metallic materials, but traditional approaches to describing microstructure capture only a small number of features. To enable data-centric approaches to materials discovery, allow efficient storage of microstructural data and assist in quality control in metals processing, we require more complete descriptors of microstructure. The concept of microstructural fingerprinting, using machine learning (ML) to develop quantitative, low-dimensional descriptors of microstructures, has recently attracted significant attention. However, it is difficult to interpret conclusions drawn by ML algorithms, which are commonly referred to as "black boxes". Here we explore variational autoencoders (VAEs), which can be trained to produce microstructural fingerprints in a continuous latent space. VAEs enable the reconstruction of images from fingerprints, allowing us to explore how key features of microstructure are encoded. We develop a VAE architecture based on ResNet18 and train it on Ti-6Al-4V optical micrographs as an example of an industrially important alloy where microstructural control is critical to performance. The latent space is explored in several ways, including by supplying interpolated and randomly perturbed fingerprints to the trained decoder and via dimensionality reduction to explore the distribution of microstructural features within the latent space of fingerprints. We show that the VAE fingerprints exhibit smooth, interpolable behaviour with stability to local perturbations, supporting their suitability as general purpose descriptors for microstructure. We also show that key properties of the microstructures are strongly correlated with position in the latent space, supporting the use of VAE fingerprints for quantitative exploration of process-structure-property relationships.