PrincipiaBlastFoam: Knowledge-Guided Multi-Agent Automation for Safety-Critical Blast Simulation

Abstract

Simulation workflows support system safety analysis, infrastructure protection, and risk-informed engineering decisions. In blast engineering, however, a case can run to completion while still violating physical assumptions through unsafe model choices, inconsistent boundary conditions, or broken cross-file dependencies. This paper presents PrincipiaBlast-Foam, a knowledge-guided multi-agent framework for safetycritical blastFoam simulation. PrincipiaBlastFoam grounds case generation in a physics-informed knowledge graph, uses graphguided retrieval to produce file-level constraints, coordinates specialized agents through an orchestrator, and applies execution and physical-validity checks before accepting a case. Experiments on a 130-task benchmark show high executability and physical validity, yielding a 91.6 percent overall success rate. Ablation and architecture studies indicate that structured domain knowledge and orchestrated state management both reduce failures in cross-file simulation setup. Validation against UFC empirical standards and urban blast gauge data provides evidence that the generated workflows preserve safety-relevant propagation behavior. Overall, dependable AI-assisted simulation benefits from explicit physical constraints combined with workflow-level verification.