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

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: simulations that run to completion while computing the wrong answer. We present Vigil, which couples deterministic science-aware triggers with multimodal LLM confirmation through a trigger-render-reason pipeline. Vigil instruments simulations with lightweight checks on derived-quantity operators and regions of interest; when a check detects a potential anomaly, the flagged data is rendered and analyzed by a multimodal model to assess scientific validity. Upon confirming invalid behavior, Vigil can terminate or restart the simulation from a checkpoint, preventing further expenditure on dark waste. We evaluate Vigil across four codes spanning numerical instability, pattern dynamics, turbulence dissipation, and molecular-dynamics failures. Vigil detects every deliberately configured failure with zero false positives and supplies an actionable semantic verdict for each alert.

Tags