# Exploring descriptors for titanium microstructure via digital fingerprints from variational autoencoders

**Authors:** Michael D. White, Gowtham Nimmal Haribabu, Jeyapriya Thimukonda Jegadeesan, Bikramjit Basu, Philip J. Withers, Chris P. Race  
**arXiv:** [2401.11967](https://arxiv.org/abs/2401.11967)  
**Date:** 2024-01-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.11967v1/

## Key points

- Microstructure is crucial in understanding and controlling metallic materials
- Traditional methods of describing microstructure are limited
- Machine learning (ML) can develop comprehensive descriptors of microstructures
- Variational autoencoders (VAEs) can generate microstructural fingerprints in a continuous latent space
- VAEs allow for the reconstruction of images from fingerprints, providing insights into how key features are encoded
- VAE architecture based on ResNet18 trained on Ti-6Al-4V optical micrographs
- VAE fingerprints exhibit smooth and interpolable behavior with stability to local perturbations
- Key properties of microstructures are strongly correlated with their position in the latent space
- VAE fingerprints can be used for quantitative exploration of process-structure-property relationships
- Potential applications include materials discovery, efficient storage of microstructural data, and quality control in metals processing.

## Summary

Microstructure is a crucial factor in understanding and controlling the properties of metallic materials. However, traditional methods of describing microstructure only capture a limited number of features. To enable data-centric approaches to materials discovery and improve quality control in metals processing, more comprehensive descriptors of microstructure are needed. Recently, the concept of microstructural fingerprinting using machine learning (ML) has gained significant attention. ML algorithms can develop quantitative, low-dimensional descriptors of microstructures. However, interpreting the conclusions drawn by these algorithms can be challenging as they are often referred to as "black boxes". In this study, the authors explore the use of variational autoencoders (VAEs) to generate microstructural fingerprints in a continuous latent space. VAEs allow for the reconstruction of images from fingerprints, providing insights into how key features of microstructure are encoded. The authors develop a VAE architecture based on ResNet18 and train it on Ti-6Al-4V optical micrographs—an industrially important alloy where precise microstructural control is critical for performance. The latent space generated by the VAE is explored through various methods, including supplying interpolated and randomly perturbed fingerprints to the trained decoder and using dimensionality reduction techniques like principal component analysis (PCA). The results demonstrate that the VAE fingerprints exhibit smooth and interpolable behavior with stability to local perturbations. This supports their suitability as general-purpose descriptors for microstructure. Furthermore, the authors show that key properties of the microstructures are strongly correlated with their position in the latent space. This finding suggests that VAE fingerprints can be used for quantitative exploration of process-structure-property relationships. Overall, this study highlights how VAEs can provide more complete descriptors of microstructure and offer insights into its key features. The use of VAE fingerprints has potential applications in materials discovery, efficient storage of microstructural data, and quality control in metals processing.