Paper Published by Energy

Title: Thermal-safety-constrained cooling energy optimization for data centers using a physics-informed graph surrogate model

Journal: Energy

DOI: https://doi.org/10.1016/j.energy.2026.142880

Abstract: The rapid growth of data center electricity demand has increased the need for energy-efficient cooling operation while maintaining local thermal safety. However, high-fidelity thermal simulation is computationally expensive, and purely data-driven surrogate models often neglect the physical topology of rack-level heat transfer and air supply paths. To address this issue, this study proposes a thermal-safety-constrained cooling energy optimization framework for non-uniformly loaded data centers. A physics-informed thermal graph is constructed to explicitly represent rack operating states, adjacent rack thermal coupling, and directed air-supply relationships from cooling units. Based on high-fidelity thermal environment simulations, a graph neural network surrogate is developed to rapidly estimate rack-level outlet temperature distributions and hotspot risks. The trained surrogate is further embedded into a differential-evolution-based optimization framework to identify energy-efficient cooling setpoints under rack outlet temperature constraints. The results show that the proposed surrogate achieves accurate rack-level temperature estimation, with a mean absolute error (MAE) of 0.379 °C and a root mean square error (RMSE) of 0.523 °C. Compared with the baseline operation, the optimized cooling strategy reduces the average local power usage effectiveness (PUE) from 1.798 to 1.603, corresponding to an energy efficiency improvement of 10.82%, an average daily electricity saving of 255.3 kWh, and a carbon emission reduction of 148.3 kg CO2. The proposed framework provides a fast and physically interpretable tool for cooling energy management and local thermal safety control in data centers.