Kalibre

Abstract

Computational fluid dynamics (CFD) model has been widely used for prototyping\ndata centers. Evolving it to high-fidelity {\\em digital twin} is desirable for\nthe management and operations of large-scale data centers. Manually calibrating\nCFD model parameters to achieve twin-class fidelity by specially trained domain\nexpert is tedious and labor-intensive. To reduce manual efforts, existing\nautomatic calibration approaches developed for various computational models\napply heuristics to search model configurations within an empirically defined\nparameter bound. However, in the context of CFD, each search step requires\nlong-lasting CFD model's iterated solving, rendering these approaches\nimpractical with increased model complexity. This paper presents Kalibre, a\nknowledge-based neural surrogate approach that performs CFD model calibration\nby iterating four key steps of i) training a neural surrogate model based on\nCFD-generated data, ii) finding the optimal parameters at the moment through\nneural surrogate retraining based on sensor-measured data, iii) configuring the\nfound parameters back to the CFD model, and iv) validating the CFD model using\nsensor-measured data as the ground truth. Thus, the parameter search is\noffloaded to the neural surrogate which is ultra-faster than CFD model's\niterated solving. To speed up the convergence of Kalibre, we integrate prior\nknowledge of the twinned data center's thermophysics into the neural surrogate\ndesign to improve its learning efficiency. With about five hours computation on\na 32-core processor, Kalibre achieves mean absolute errors (MAEs) of $0.81^o$C\nand $0.75^o$C in calibrating two CFD models for two production data halls\nhosting thousands of servers each while requires fewer CFD solving processes\nthan existing baseline approaches.\n