Workshops Archive

PULSE: Parallel Uncertainty-guided Landscape Sampling Engine for In-Situ Molecular Dynamics


Workshop: ISAV26: In Situ AI, Analysis and Visualization

Authors: Harshita Sahni (University of New Mexico); Silvina Caino-Lores (University of Rennes); Jack Marquez (University of Tennessee); Ewa Deelman (University of Southern California); Michel Cuendet (Swiss Institute of Bioinformatics); Harel Weinstein (Weill Cornell Medical Center); Michela Taufer (University of Tennessee, Knoxville); and Trilce Estrada (University of New Mexico)

Abstract: Molecular Dynamics (MD) simulations generate massive streams of high-dimensional data. However, biologically significant events are rare, transient, and not known ahead of time. Modern MD workflows execute many trajectories concurrently, but most adaptive sampling methods analyze trajectories in offline batches, delaying feedback and wasting computational resources on redundant exploration. We present PULSE, a scalable in-situ framework that closes the feedback loop between simulation and analysis through a synchronized producer-consumer runtime. PULSE performs online conformational state identification, detects previously unexplained regions of the conformational landscape, and adaptively redirects computational effort toward these regions during execution. We evaluate PULSE on two protein systems, Villin Headpiece and beta-lactoglobulin. Compared to conventional adaptive sampling, PULSE drives 70.77% and 37.17% of newly sampled conformations, respectively, into regions that the baseline never explored. We demonstrate end-to-end execution on eight GPUs and show that PULSE maintains nearly constant coordination overhead across concurrency levels up to 32 trajectories.


Back to ISAV26: In Situ AI, Analysis and Visualization Archive Listing

Back to Full Workshop Archive Listing