Multi-Agent Orchestration of Optimization Workflows for Richtmyer–Meshkov Instability Studies

  • Fernandez-Godino, María Giselle (Lawrence Livermore National Laboratory)
  • Belof, Jonathan (Lawrence Livermore National Laboratory)
  • Brown, Nathan (Sandia National Laboratories)
  • Gunnarson, Brian (Lawrence Livermore National Laboratory)
  • Hammond, Ruth (Florida International University)
  • Menon, Harshitha (Lawrence Livermore National Laboratory)
  • Li, Yin-Wai (Lawrence Livermore National Laboratory)
  • Shachar, Meir (Lawrence Livermore National Laboratory)
  • Sterbetz, Dane (Lawrence Livermore National Laboratory)
  • Thevenin, Sebastien (Lawrence Livermore National Laboratory)
  • White, Daniel (Lawrence Livermore National Laboratory)
  • Jekel, Charles (Lawrence Livermore National Laboratory)
  • Schi;;, William (Lawrence Livermore National Laboratory)

Please login to view abstract download link

Scientific design problems often involve large ensembles of simulations, resource scheduling, and performance evaluation. We introduce a multi‑agent [1] design assistant (MADA) based on the Model Context Protocol (MCP) [2] that orchestrates these tasks via specialized agents. A large‑language‑model orchestrator interprets high‑level objectives and coordinates simulation, geometry, surrogate and scheduling agents. The simulation agent generates parameter sweeps from template scripts and runs them on high‑performance computing systems, while the scheduler agent manages job submissions and monitors progress. The geometry agent supplies meshes and initial conditions, and the surrogate agent uses machine‑learning models to propose new designs, reducing reliance on expensive simulations. By decoupling these responsibilities, the orchestrator assembles parameterized runs, delegates execution, and, on completion, triggers an analysis module that computes quantities of interest, evaluates scalar objectives, and reports the optimal design along with summary data and optional visualizations. Communication through MCP tools confers modularity: new codes, job managers, and analysis pipelines can be integrated without modifying the orchestrator. Demonstrated on shock‑physics problems such as Richtmyer–Meshkov instability suppression [3], this architecture provides a flexible foundation for automated, closed‑loop optimization and design exploration across diverse scientific domains.