[CS] Mike Tynes MS Presentation/Aug 12, 2025

[via cs]

This is an announcement of Mike Tynes's MS Presentation
===============================================
Candidate: Mike Tynes

Date: Tuesday, August 12, 2025

Time:  1 pm CST

Location: JCL 298

Remote Location: https://uchicago.zoom.us/j/91533660351?pwd=bO0xgZOdFXHXrSm8LRensxxII2Rlo2.1

Title: On-the-fly training of machine-learned surrogates for dynamical simulations

Abstract: Training and deploying inexpensive machine-learned (ML) surrogates for computationally expensive subroutines "on-the-fly" (OTF) during physical simulations offers both potential advantages and unique drawbacks. In OTF learning, an ML surrogate is trained to replace a target subroutine as a simulation evolves. The advantages of OTF learning include reducing simulation error and model training costs, while the weaknesses include the possibility of introducing artifacts when adaptively updating the model which drives the physics of the simulation. This work extends an existing approach called Proxima, which uses a control system to train an OTF surrogate such that the average error over simulation steps meets a user-specified error bound. The existing Proxima approach is appropriate for non-dynamical state space sampling approaches (such as Monte Carlo methods), but our results show that it can introduce artifacts when simulations are explicitly evolved over time according to equations of motion. To solve this problem and extend Proxima to dynamic simulations, we introduce a "blending" procedure, called Proxima+Blend, that removes discontinuities when transitioning between the expensive subroutine and surrogate by evolving the simulation according to a mixture of the forces obtained from both surrogate and target functions. The mixing coefficient varies smoothly between zero and one over time according to both uncertainty quantification of the ML surrogate and observed errors when the target subroutine data is available. We show that while the original Proxima implementation can shorten simulation runtime and accurately capture some macro-scale properties in molecular dynamics simulations, it introduces unphysical dynamics at short time and length scales and for some dynamical properties it can introduce as much as 80% error. Meanwhile, our new approach Proxima+Blend delivers a 1.5x speedup over use of the target subroutine while estimating measuring these same properties within 5% error. Our implementation of Proxima+Blend can be deployed by simply replacing the existing subroutine with a wrapper that includes a machine learning approach for surrogate training, along with specifying control and uncertainty signals for the simulation of interest.

Advisors: Ian Foster and Kyle Chard

Committee Members: Kyle Chard, Logan Ward and Ian Foster