"Vision-Based Machine Learning Methods for Microstructure Synthesis" 

Aidan Pantoya, MSE PhD Candidate 

Advisor: Professor Alejandro Strachan

ABSTRACT

The growth of computational power and the development of new machine learning (ML) methods are driving a transformative era in materials science. This confluence has led to innovative vision-based ML applications, such as the generative modeling of digital microstructures, a domain where traditional methods often face significant challenges. Experimental techniques are limited in their ability to capture and track microstructural influences across high spatial and temporal resolution. Physics-based simulations, such as molecular dynamics (MD), provide detailed insights but are restricted to small length and time scales due to their computational cost. Continuum-level simulations reduce the computational demand but introduce greater approximations, losing the detail maintained by atomistic simulations. To capture the diverse and broad range of microstructural influences, computational simulation techniques require large amounts of microstructure data obtainable only through high throughput frameworks. Vision-based neural networks offer a promising solution by enabling the efficient synthesis of realistic microstructures that can subsequently be evaluated through simulation. This preliminary report presents an investigation of vision-based ML methods, including convolution-based variational autoencoders, generative adversarial networks, and diffusion models. Each network has the potential to enhance our understanding of microstructure-property relationships by generating diverse and physically plausible microstructures that can serve as high-quality inputs to property prediction simulations.