MSE Seminar: Jenna Pope

Event interval: Single day event
Campus room: IEB G106
Accessibility Contact: Matthew Yankowitz, myank@uw.edu
Event Types: Lectures/Seminars

Title: BUQ4MLIP: Uncertainty Quantification for Machine Learning Interatomic Potentials

Abstract: Machine learning interatomic potentials (MLIPs) are transforming atomistic simulations by providing ab initio-level accuracy at orders-of-magnitude faster speeds. Universal MLIPs, also called foundation models, are trained across large, chemically diverse datasets with the goal of broad transferability to new atomic environments. However, the blackbox nature of MLIPs obscures the relationship between atomic environments and predicted outputs, making it difficult to discern when the model is operating within its domain of validity and when its outputs may be unreliable. Uncertainty quantification (UQ) provides interpretable metrics that indicate the confidence or reliability of MLIP predictions. This talk will discuss UQ approaches using quantile regression to produce confidence intervals around the predicted output that can be used to assess the trustworthiness of MLIP-driven simulations.

Bio: Dr. Jenna Pope (who publishes under the name Jenna A. Bilbrey) is a Data Scientist at Pacific Northwest National Laboratory, working within the National Security Directorate. Her research bridges computational chemistry, materials science, and AI/ML, with a focus on applying deep learning and uncertainty quantification methods to chemical and materials modeling. She holds a PhD in computational chemistry from the University of Georgia and a BS in chemistry from the University of West Florida. Her projects are highly interdisciplinary and involve close collaboration with both experimentalists and theoreticians. Over the course of her career, she has published on topics such as neural network potentials, active learning for materials simulations, and graph-component methods for defect analysis.