Reducio: model reduction for data center predictive digital twins via physics-guided machine learning

The digital twin, as a digital counterpart of a physical entity, has shown great potential in data center prototyping and predictive thermal management. In this regard, Computational Fluid Dynamics/Heat Transfer (CFD/HT) models have been widely adopted. However, the computing time of the CFD/HT simulation is prohibitively long in practice. The Proper Orthogonal Decomposition (POD) has been explored to approximate the CFD/HT simulation by a linear combination of POD modes and coefficients. Existing approaches to calculating the POD coefficients use either the black-box interpolation or the simplified physical model, leading to unsatisfactory generalization ability. To advance existing approaches, this paper proposes Reducio, a physics-guided model reduction approach based on the POD to predict the temperature field by following two key phases of i) offline POD modes calculation and coefficients interpolation and ii) online coefficients extrapolation supervised by the principle of energy balance. To extrapolate the coefficients with limited training data, we adopt the Gaussian Process (GP) model to learn a nonlinear map between the boundary conditions and POD coefficients. We conduct two case studies in two data centers with different scales. Evaluation results in the edge data center show that Reducio achieves sub-1°C mean absolute error (MAE) in temperature field prediction compared with the CFD/HT simulation result, outperforming the existing method based on the simplified physical model by 1.5 °C. When evaluating in the industry-grade hyper-scale data center with the sensor measurements, around 1°C temperature prediction MAE is observed. Furthermore, Reducio can predict the full-fledged temperature field in real-time, making it a strong candidate for building data center predictive digital twins.