Toward Agentic HPC: Serving and Evaluating LLM-Powered Agents for Scientific Applications On Leadership-Class Platforms

HPC simulations are essential to scientific discovery but remain difficult to fully automate due to complex workflows, long queue times, and expensive evaluations. Recent advances in large language models (LLMs) make intelligent automation via agentic modeling and simulation (ModSim) increasingly viable, but integrating LLM agents into HPC environments introduces significant systems and scalability challenges. We propose and demonstrate an agentic HPC framework that enables LLM-driven optimization of scientific ModSim through a persistent model service, scalable agent orchestration, and a generic interface to HPC applications and schedulers. Our design explicitly accounts for HPC constraints, including batch scheduling, failures, and heterogeneous accelerator resources.
We evaluate the framework on two representative scientific applications—materials design and weather modeling—deployed on leadership-class systems at the Oak Ridge Leadership Computing Facility. Across multiple frontier LLMs, the agentic approach achieves up to 1.7x faster time-to-solution, compared to evolutionary algorithms, while reducing computational cost. We further demonstrate scalable, multi-node LLM serving on AMD and NVIDIA GPUs, highlighting the practical feasibility of agentic workflows on modern supercomputers. These results suggest that agentic AI can serve as a practical and effective abstraction for automating HPC simulations, improving both scientific productivity and resource efficiency.

---

Contributors:

• Mijanur Palash
• Feiyi Wang
• Sarp Oral
• Mallikarjun Shankar