Fusion Intelligence for Digital Twinning AI Data Centers: A Synergistic GenAI-PhyAI Approach

TL;DR

Fusion Intelligence presents a dual-agent framework that unites GenAI’s generic, language-driven generation with PhyAI’s physics-grounded optimization to create and refine digital twins for AI-dedicated data centers (AIDCs). The outer loop uses GenAI to generate semantic digital-twin configurations, while the inner loop employs PhyAI to calibrate parameters under physical laws and live data, forming a closed-loop workflow. Case studies demonstrate significant improvements: physics-aware equipment selection yields a PUE of 1.25 versus 1.35 from heuristic methods, and generative heat-exchanger modeling reduces mean percentage error to 2.2% from 6.3% in expert models. These results indicate that AI-native Fusion Intelligence can accelerate digital transformation of mission-critical infrastructures by delivering reliable, scalable, and physically plausible digital twins across the AIDC lifecycle.

Abstract

The explosion in artificial intelligence (AI) applications is pushing the development of AI-dedicated data centers (AIDCs), creating management challenges that traditional methods and standalone AI solutions struggle to address. While digital twins are beneficial for AI-based design validation and operational optimization, current AI methods for their creation face limitations. Specifically, physical AI (PhyAI) aims to capture the underlying physical laws, which demands extensive, case-specific customization, and generative AI (GenAI) can produce inaccurate or hallucinated results. We propose Fusion Intelligence, a novel framework synergizing GenAI's automation with PhyAI's domain grounding. In this dual-agent collaboration, GenAI interprets natural language prompts to generate tokenized AIDC digital twins. Subsequently, PhyAI optimizes these generated twins by enforcing physical constraints and assimilating real-time data. Case studies demonstrate the advantages of our framework in automating the creation and validation of AIDC digital twins. These twins deliver predictive analytics to support power usage effectiveness (PUE) optimization in the design stage. With operational data collected, the digital twin accuracy is further improved compared with pure physics-based models developed by human experts. Fusion Intelligence offers a promising pathway to accelerate digital transformation. It enables more reliable and efficient AI-driven digital transformation for a broad range of mission-critical infrastructures.

Figures (7)

• Figure 1: Growth of computational performance demand over the past four decades. Moore’s law is broken after entering the large model era zhu2023intelligent. Such growth introduces new challenges for AIDC infrastructure management.
• Figure 2: The overview of the AIDC system that consists of the power supply, computing, and cooling systems. Compared to traditional DC, AIDC is usually equipped with a liquid cooling system to remove the heat generated by high-power-density racks. The interconnected system presents multiple modeling challenges such as large-scale layout construction, data availability, data quality, multiphysics, and multiscale nature.
• Figure 3: Intelligence evolution for DC management. Human intelligence relies heavily on human expertise and manual workflows; Machine intelligence introduces efficiency gains but faces challenges in generalizability and physical consistency; Fusion intelligence integrates GenAI’s automation with PhyAI’s physical grounding, creating a closed-loop framework for scalable and reliable DC operations.
• Figure 4: The fusion intelligence operates as follows: 1) The GenAI samples tokenized digital twin configurations and codes from prompts and the feedback from the PhyAI optimization. 2) These samples are then simulated using a time series dataset. 3) The simulation results are evaluated and organized into numerical errors and semantic descriptions, providing feedback for PhyAI and GenAI, respectively. 4) The PhyAI optimizes model parameters and 5) GenAI reflects the semantic results to refine the generation.
• Figure 5: The proposed system architecture integrates the capabilities of GenAI and PhyAI agents into a fusion intelligence framework designed for AIDC lifecycle management and relevant applications.

Theorems & Definitions (2)

• Definition 1: Agents
• Definition 2: Fusion Intelligence