[CS] Trenton Wesley MS Presentation/Dec 16, 2025

[via cs]

This is an announcement of Trenton Wesley's MS Presentation
===============================================
Candidate: Trenton Wesley

Date: Tuesday, December 16, 2025

Time:  4 pm CST

Remote Location: https://uchicago.zoom.us/j/93672609001?pwd=DFFZgLGQKmRo3KY8J8ZAn0mLw6dSbG.1  Meeting ID: 936 7260 9001 Passcode: 449184

Location: JCL 298

Title: Delta: Uncertainty-Aware Control for On-the-Fly Surrogates in Atomistic Simulations

Abstract: On-the-fly, machine-learned surrogate models can often safely replace expensive physics models during atomistic simulations, but a control strategy is required to decide when to use the surrogate to balance speedup and accuracy. In prior work, there are both static and dynamic control schemes for deciding when to use a surrogate model at each step of a simulation through the usage of an uncertainty signal typically from the simulation. However, these approaches are limited by (1) ad-hoc uncertainty signals that have an unknown relationship with surrogate error and (2) by a failure to connect iteration-level surrogate error to the final simulation error, which in practice forces problem-specific tuning to find a valid surrogate error bound. We address these limitations by using a measure of epistemic and aleatory uncertainty from the surrogate model itself and by formally relating per-step surrogate error to the final simulation-level error, yielding a noticeably improved tradeoff between speed and accuracy. We introduce the Desired-Error Learning Tolerance Adapter (DELTA), a control scheme that uses ensemble variance to obtain a more informative uncertainty measure and analytically derives a stepwise surrogate-error threshold from a user-specified bound on the observable. In our experiments, ensemble-based uncertainty features significantly improve error prediction and controller decisions, eliminating the need for problem-specific tuning and enabling safer, more generalizable acceleration of atomistic simulations.

Advisors: Hank Hoffmann

Committee Members: Hank Hoffmann, Ian Foster, and Yuliana Zamora