Using a simulated HPC cluster to evaluate candidate solvers, we discovered a linear solver for FLASH magnetic diffusion that reduces AMG setup time by approximately 49% and achieves an overall speedup of 1.52×.
Evaluating thousands of AI-generated linear solvers requires consistent hardware and sub-millisecond timing sensitivity. We built a custom sandbox infrastructure on bare-metal AWS EC2 instances to reduce timing noise and isolate solver evaluations from noisy neighbors.
We built a system to discover faster linear solvers for magnetic-diffusion systems arising in FLASH. To ground the discovery process, we assembled an initial dataset of linear systems from parameter variations of AlWire problems 1 and 2 and the ZPinch example, together with two MagDiff unit tests: the implicit-AMR test and the edge test. We achieved a 1.546× speedup (54.6%) over baseline.
Faster solvers for 1D Linear Advection and 1D Viscous Burgers
