LC-Opt: RL and Agentic AI for Liquid Cooling Optimization

Avisek Naug and co-authors present LC-Opt, a Modelica-based benchmark environment for reinforcement-learning and agentic-AI control of end-to-end liquid cooling in high-density data centers (digital twin based on Oak Ridge National Laboratory’s Frontier cooling system).

• Main announcement: The paper introduces LC-Opt, a sustainable liquid cooling benchmark that provides Modelica-based end-to-end models (site-level cooling towers to data center cabinets and server blade groups) built on a high-fidelity digital twin of Oak Ridge National Lab’s Frontier Supercomputer cooling system; it exposes controls (liquid supply temperature, flow rate, granular valve actuation, cooling tower setpoints) via a Gymnasium interface and supports optional components like a Heat Recovery Unit (HRU).
• Background and details: The authors benchmark centralized and decentralized multi-agent RL, perform policy distillation into decision and regression trees for interpretability, and explore LLM-based methods and an agentic mesh architecture for natural-language explanations of control actions. Submitted to NeurIPS 2025 (submission date: 31 Oct 2025).