Toward Automated Calibration of Data Center Digital Twins: A Neural Surrogate Approach

Evolving the computational fluid dynamics (CFD) model to high fidelity digital twin is desirable for industrial data center management. However, existing CFD model calibration approaches to improve the model accuracy require either excessive manual tuning or intensive computation, rendering them non-scalable with system size and complexity. This paper presents a surrogate-based approach to automate the calibration of CFD models built for industrial data centers. Specifically, a knowledge-based graph neural net (GNN) is trained to approximate a CFD model as a surrogate model that captures the key thermal variables and their causal relationships in a given data hall. By integrating prior knowledge as constraints, the GNN has reduced demand on the amount of training data. After rounds of the training processes, the neural surrogate can recommend the optimal configurations for the CFD model parameters that are hard to obtain, such that the temperatures predicted by the CFD are most consistent with the actual measurements. Experiments of applying the proposed approach to calibrate two CFD models built for two production data halls hosting thousands of servers achieve temperature prediction errors of $0.81^\circ$C and $0.75^\circ$C with about $30$ hours of computation on a quad-core virtual machine in the cloud.