Agentic LLM-Based Digital Twin for Liquid Cooling

This section describes a modular agent-based architecture enhanced with Large Language Models (LLMs) to enable explainable control in liquid-cooled data centers. The system is built on functional agents that communicate via a centralized Message Bus.

Agent Types

• LLM/LLM-supported Agents (purple): Handle reasoning, planning, and control decisions.

• Non-LLM Agents (blue): Handle system-specific monitoring, control, and interfacing.

• Tools (green): Provide analytics, modeling, and visualization.

Main Components

• Orchestration: Manages agent lifecycle, communication, and error handling.

• Agent Monitor: Tracks performance and suggests optimizations.

• Maintenance: Detects anomalies and enables predictive maintenance.

• Math Toolbox: Offers modeling tools, statistical reasoning, and uncertainty analysis.

• Configuration: Handles system parameters, logic, and hyperparameter tuning.

• Visualization: Supports explanations, dashboards, and trend displays.

• Control: Issues real-time actions, combines RL and LLM-based policies.

• Sensor: Interfaces with the physical system and digital twin.

• LLM Control Agents: Issue adaptive actions for cooling elements based on historical and real-time data.

• Liquid Cooling System: Uses a CDU and cabinet valves for precise thermal control, based on sensor feedback (T_in, T_out).

Use this architecture to combine physical system modeling, real-time control, and explainable AI for advanced cooling management.

• LLM Agentic Playground
  • Configurable Reasoning for Trade-Off Testing
  • Drag-and-Drop Agent Design for Scalable Control
  • Stress Injection for Reasoning Evaluation
  • Real-Time Monitoring
  • Explainability Tools
  • Safety and Guardrails
• LLM Agentic Explainable Control
  • Executive Summary Comparison
  • Cooling Assessment
  • Short-Term Recommendation
  • Qwen 3-8B: Short Explanation
  • Qwen 3-8B: Detailed Explanation