Technical Papers Archive

Trigger, Render, Reason: Multimodal LLM Agents for Science-Aware Scientific Simulation Monitoring


Authors: Hua Xu and Jaime Cernuda (Illinois Institute of Technology), Ana Gainaru and Scott Klasky (Oak Ridge National Laboratory (ORNL)), and Anthony Kougkas and Xian-He Sun (Illinois Institute of Technology)

Abstract: Scientific simulations on leadership-class HPC systems routinely waste millions of core-hours producing invalid results, yet no tool monitors whether a running simulation computes valid science. We call this blind spot dark waste. Detecting it requires physics-aware judgment at every timestep, which neither humans nor system-level monitors supply. Multimodal LLM agents could, but they cannot ingest scientific data: no field fits in a context window. What an agent can ingest is a rendering, a lossy encoding that discards precision, preserves structure, and stays a fixed size regardless of the field behind it. Because a fixed rendering answers only the question it was made for, we give the agent control of the visualization. Vigil instantiates this as a trigger-render-reason pipeline: cheap deterministic triggers watch every timestep, and only flagged states are rendered and reasoned over. Across four codes, Vigil detects every parameter-induced failure and dismisses a trigger fired on a healthy run.


Presentation: file

Back to Technical Papers Archive Listing