MADA study automates HPC design iterations for fusion simulations

Researchers presented MADA, an LLM-powered multi-agent framework that coordinates simulation jobs, geometry generation and inverse design workflows on high-performance computing systems. In tests focused on suppressing Richtmyer-Meshkov instability, the system iteratively refined designs using hydrodynamics simulations and a pre-trained machine-learning surrogate. The authors report reduced manual workflow setup and a reusable approach to coupling reasoning, specialized tools and simulation-driven design exploration.