Speaker
Description
Approximately 70% of the matter content of our Universe consists of dark matter, yet its fundamental nature remains unknown. Among the many proposed candidates, the axion stands out as particularly compelling because it not only provides a viable explanation for dark matter but also offers an elegant solution to the strong CP problem, one of the most persistent puzzles in the Standard Model of particle physics. As a result, a broad experimental program has emerged worldwide to search for axions using a variety of detection strategies.
A major challenge faced by many of these experiments is that the axion mass is not known. Because detection relies on resonant enhancement, experiments must scan over a wide range of possible masses, making searches time-consuming and costly. Reliable theoretical predictions of the axion mass would dramatically improve the efficiency of these efforts. However, obtaining such predictions requires understanding the highly non-linear dynamics of the axion field in the early Universe. This regime can only be studied through large-scale numerical simulations of coupled scalar fields.
Conceptually, these simulations are simple and only involve solving classical field equations on a three-dimensional mesh. In practice, they are extremely demanding. The formation of topological defects known as axion strings introduces a severe separation of physical scales. Accurately resolving string cores while simultaneously capturing large cosmological volumes requires enormous dynamic range. Uniform-grid simulations have reached trillions of cells, yet still struggle to capture the full evolution with controlled systematics.
To address this challenge, we employ adaptive mesh refinement (AMR), which dynamically concentrates resolution only where it is required. We developed sledgehamr, a simulation framework built on AMReX, to make large-scale scalar-field simulations with AMR efficient and accessible on both GPU and CPU systems. The code provides a flexible and modular infrastructure originally designed for axion string networks but readily extensible to other problems, such as gravitational waves sourced by scalar-field dynamics.
Using this framework, we performed the largest three-dimensional simulation of axion strings to date on the NERSC Perlmutter system (see https://tinyurl.com/AxionStringsAMR for an animation). The simulation follows the evolution of the axion field during the first microseconds after the Big Bang, when strings form and dominate the dynamics. It consumed approximately 25 million CPU core-hours and produced roughly 500 TB of output data. By leveraging AMR, we achieved a dynamic range about 3000 times larger than previous studies. Reaching a comparable resolution with a static grid would have required more than $10^{16}$ cells and would be computationally infeasible.
This unprecedented scale substantially reduces systematic uncertainties in string evolution and axion production, enabling the most robust prediction to date of the axion dark-matter mass and helping to narrow the search space for experiments. More broadly, our results demonstrate that adaptive mesh refinement is essential for tackling the extreme multiscale challenges that arise in early-Universe field dynamics and highlight AMReX as a powerful platform for next-generation HPC cosmology simulations.
