Case Study: Argonne National Laboratory achieves scalable, higher-fidelity quantum molecular dynamics with Altair PBS Professional

Argonne systems used to smooth out Quantum molecular dynamics simulations

Argonne National Laboratory, a U.S. Department of Energy multidisciplinary research center with leadership-class systems like the 34-petaflop Polaris and upcoming Aurora, faced a fidelity-scaling challenge: state-of-the-art neural-network quantum molecular dynamics (NNQMD) models such as Allegro became unstable and failed at large scale. Altair is referenced as the vendor whose PBS Professional workload manager orchestrates GPU workloads on Polaris, supporting the high-performance environment needed for these demanding simulations.

To solve the problem, researchers combined the Allegro model with sharpness-aware minimization to create “Allegro-Legato” and ran large GPU-accelerated jobs on Polaris with orchestration provided by Altair PBS Professional. The Allegro-Legato approach preserved Allegro’s accuracy and speed while greatly improving robustness and scaling — force outliers remain nearly constant and time-to-failure increases by nearly a factor of two as problem size grows — enabling larger, longer NNQMD simulations on leadership-scale and exascale platforms.

Argonne National Laboratory

Hikaru Ibayashi

Argonne National Laboratory