CAPER: Dual-level physics-data fusion with modular metamodels for reliable generalization in predictive digital twins

Digital twins enable Artificial Intelligence (AI)-driven decision-making in mission-critical infrastructures such as data centers by leveraging advanced predictive capabilities. However, deploying predictive digital twins in a data center presents significant challenges, particularly in ensuring real-time updates and reliable dynamic predictions. Existing approaches capable of capturing complex system characteristics, ensuring predictive accuracy on unseen datasets, and balancing computational overhead remain lacking. This paper proposes a dual-level physics-data fusion framework that synergistically fuses physics-based and data-driven methods. The framework employs system decomposition and modular metamodeling to reduce sampling requirements and improve complex parameter search efficiency. Experimental validation demonstrates that the Coefficient of Variation of the Root-Mean-Square Error (CV-RMSE) remains below 15 % for most features, outperforming baseline methods where individual feature errors range from 40 % to 100 %. The single-time calibration completes within the typical 15-minute sampling interval in cooling systems, thereby enabling real-time updates. By seamlessly integrating the extrapolation capabilities of physics-based models with the adaptability of Machine Learning (ML), the framework achieves high predictive accuracy in both interpolation and extrapolation tasks. The adaptation process further reduces the Mean Absolute Error (MAE) by up to 93 % compared to the Baseline. Furthermore, the fusion-based system dynamics model effectively simulates multi-stage heat transfer processes, integrating system-level thermodynamic interactions with module-level metamodels. This framework demonstrates significant potential to enhance digital twin applications, optimize data center energy efficiency, and address broader engineering challenges.