Autonomous science campaigns¶
A systematic study running real physics, chemistry, and engineering campaigns end-to-end across ALCF, NERSC, and OLCF systems — not to produce a single result, but to examine how well an autonomous agent handles failure, self-corrects, and knows when a result is trustworthy, not just whether a job finished. Across the full study, agents self-corrected roughly 49 of 65 recorded iteration failures with zero human intervention.
-
Finding argon's melting point via molecular dynamics
A 7-temperature LAMMPS sweep locates the solid-to-liquid transition within 4% of the literature value.
-
Silicon's electronic band structure from first principles
A four-step Quantum ESPRESSO DFT workflow computes silicon's band gap in 24 seconds on one GPU node.
-
Trp-cage mini-protein thermal stability across three temperatures
Three GROMACS runs show a mini-protein's structure destabilizing steadily as temperature climbs.
-
Water dimer binding energy — DFT vs. quantum Monte Carlo
A clean DFT result, plus an honestly reported handoff failure when refining it with quantum Monte Carlo.
-
Turbulent pipe flow across two facilities: Polaris vs. Aurora
The same CFD simulation on NVIDIA and Intel GPUs, five build iterations apart, within 13% throughput of each other.
-
Reproducing the same GPU bug on two different supercomputers
A cosmology code's GPU crash reproduced identically on Polaris and Perlmutter, confirming a real code defect.
-
GROMACS water box: GPU (Polaris) vs. CPU-only (Frontier)
A cross-facility performance baseline that later fed a follow-on GPU-porting effort on Frontier.
-
Opening silicon's band gap the right way: G0W0 many-body corrections
A many-body GW correction pushes silicon's DFT gap toward its true value, as theory predicts.
-
Breaking a nitrogen molecule apart, three ways, across three supercomputers
DMRG matches exact full configuration interaction to nine decimal places, verified on three systems.
-
Copper's melting point via a heating-ramp simulation
A continuous heating ramp finds copper's melting discontinuity, showing the expected superheating artifact.
-
Pushing turbulent pipe flow to a higher Reynolds number
The same spectral-element method holds stable and performant at nearly 4x the turbulence intensity.
-
Running four molecular dynamics replicas at once
Four independent GROMACS replicas in one job, each on its own GPU, with negligible parallel overhead.
-
Silicon carbide's vibrational spectrum from first principles
DFPT phonon frequencies land within 1% of experiment, with symmetry checks passing cleanly.
-
Verifying HPC applications actually run
A coverage sweep across ~38 applications and 9 systems, checking for a physically sensible result, not just a clean exit.