Announcement Detail
Thursday, October 23, 2025
3:00 PM EDT
Join via Zoom: https://us06web.zoom.us/j/
USACM UQ Virtual Seminar
Heteroscedastic Gaussian Process Regression and Active Learning for Microstructure–Property Predictions
Speaker
Lori Graham-Brady, Johns Hopkins University
Abstract
Understanding the relationship between microstructure and material performance remains a fundamental challenge in computational materials science, particularly for complex behaviors such as effective property prediction, energetic material response, and spall behavior. Recent advances in machine learning offer powerful new tools to address this challenge. In particular, Gaussian Process Regression (GPR) provides a flexible, non-parametric modeling framework that can predict material performance from microstructural descriptors while also quantifying uncertainty. This seminar will explore the use of GPR in conjunction with active learning (AL) to efficiently sample microstructural design spaces and guide materials discovery. The first part of the talk presents a one-dimensional study that incorporates heteroscedastic (input-dependent) aleatoric uncertainty into the GPR framework, which is a key feature in many mechanics problems where predictive variability changes across the microstructural input parameter space. Building on this, the seminar introduces a batch active learning strategy designed to accommodate such heteroscedasticity, and evaluates how different sampling strategies influence learning efficiency and predictive performance in relatively simple one- and two-dimensional functions, as well as in effective property predictions for composite materials. The final part of the talk demonstrates applications of this approach to more complex material systems, including energetic materials and ongoing efforts to model spall behavior. Collectively, these results highlight both the opportunities and challenges of using uncertainty-aware surrogate models to bridge the gap between microstructure and performance in data-scarce materials applications.
Upcoming webinar speakers:
Thursday, November 13: Drew P. Kouri, Sandia National Laboratories
Thursday, December 11: Ionut-Gabriel Farcas, Virginia Tech