2025 International Conference on Sensing, Measurement & Data Analytics in the era of Artificial Intelligence (ICSMD 2025)

Efficient and reliable cold source supply is vital for large-scale supercomputing systems, especially during batch computing or training, where temperature fluctuations directly affect throughput and hardware lifespan. However, conventional cooling strategies often struggle to adapt to the dynamic and complex thermal loads in modern high-performance computing (HPC), resulting in energy inefficiency and potential hardware failures. This paper introduces DCAIM-GPT—a mechanism-knowledge hybrid framework that integrates high-fidelity digital twins with a multi-agent decision engine empowered by enhanced Retrieval-Augmented Generation (RAG) technology. This framework autonomously optimizes critical operational parameters—including but not limited to the number of operating pumps, pump operating frequencies, fan speeds, and cooling tower hierarchical operation modes—under fluctuating computational workloads and outdoor temperature/humidity conditions. It incorporates a large-model-driven multi-agent iterative mechanism to determine optimal parameters while ensuring the supercomputer's cooling source remains within safe operating limits. Experiments in digital twin environments demonstrate that DCAIM-GPT consistently maintains cooling safety and enhances energy efficiency under both dynamic and extreme conditions, evidenced by stable PUE metrics and improved coefficient of performance (COP). This research provides a scalable, universal solution for large-model-empowered efficient and reliable supercomputing operations, laying the foundation for broader vertical industrial large-model applications across the supercomputing domain.