CFD Surrogates for Data Center Sustainability Using 3D U-Net Convolutional Neural Network

Data centers often suffer from energy inefficiencies caused by suboptimal heat flow, cooling imbalances, and temperature disparities. Although Thermal Computational Fluid Dynamics (CFD) can model these challenges, its computational demand often impedes timely, adaptive energy-saving measures. Addressing this, we introduce an efficient machine learning CFD surrogate model, DCCFD, implemented by a 3D Convolutional Neural Network (CNN) trained on the CFD or IOT-generated data. DCCFD has an innovative input channel design that captures the data center’s 3D geometry, power usage, airflow dynamics, and AC unit set points in three dimensions, enabling 3D CNN models like U-Net to accurately predict and generate the heat map. DCCFD is orders of magnitude faster than the CFD model with high accuracy. This enables real-time control for dynamic reallocation of computational load and cooling system adjustments, yielding significant energy savings. Beyond real-time adjustments, our approach facilitates proactive optimization during the DC installation phase, minimizing potential hot spots and maximizing equipment lifespan. So, DCCFD contributes to carbon footprint reduction and sustainability by maximizing cooling efficiency, reducing energy consumption, and increasing performance while extending the life of servers.