Kalibre: Knowledge-based Neural Surrogate Model Calibration for Data Center Digital Twins

Overview

This paper, by Ruihang Wang and colleagues, was published in 2020 at ACM BuildSys (pages 200-209; DOI 10.1145/3408308.3427982, arXiv:2001.10681). It tackles the step that usually breaks data-hall CFD twins: parameters that are hard to measure, especially per-server airflow. This airflow depends on the server's internal fan control and is rarely in vendor specifications. The record rests only on the abstract and the paper's headline results.

The method is as follows: a neural surrogate is trained on CFD runs, the known thermophysics of the hall is embedded in the network, and parameters are searched on the surrogate against cold- and hot-aisle sensor readings. The best parameters are written back to the CFD model and re-checked against measurements, iterating until converged. According to the paper, manually built CFD models can give temperature prediction errors up to 5 C, citing prior work.

The results come from two production halls of hundreds of square metres hosting thousands of servers. The calibrated CFD achieved mean absolute errors (MAE) of 0.81 C and 0.75 C at tens of cold and hot aisle sensor positions, after about 5 hours on a 32-core VM. With the same compute budget, heuristic search and a plain neural surrogate reached MAEs of about 1.5-4 C, and manual calibration by a domain expert gave 1.32 C and 1.1 C (63% and 46% higher than Kalibre).

The paper also notes that sensors in the aisles mainly reflect nearby CRACs, and that inlet temperatures are often required to be between 15 C and 27 C. As a limit, the results are confined to two halls, and generalisation of the method to other hall types must be verified separately. It still offers a practical, measurable workflow for post-commissioning calibration of digital twins and a realistic accuracy target.

Key points

1. Manually built CFD models can give temperature prediction errors up to 5 C (citing prior work), largely because per-server airflow is unknown and hard to measure at scale.
2. Kalibre loop: (1) train a surrogate on CFD data, (2) search parameters by retraining the surrogate on sensor data, (3) write parameters back to CFD, (4) validate CFD against sensor ground truth.
3. For two production halls of hundreds of square metres hosting thousands of servers, calibrated CFD achieved MAEs of 0.81 C and 0.75 C at tens of sensor positions, after about 5 hours on a 32-core VM.
4. With the same compute budget, heuristic search and a plain neural surrogate reached MAEs of about 1.5-4 C, and manual expert calibration gave 1.32 C and 1.1 C (63% and 46% higher than Kalibre).
5. Aisle sensors mainly reflect nearby CRACs, and inlet temperatures are often required to be between 15 C and 27 C.
6. The method embeds physical knowledge in the neural network; the aim is to automate calibration.
7. DOI 10.1145/3408308.3427982; arXiv:2001.10681.

Using it in design

• Include a calibration workflow in the hand-over to operations: import aisle sensor data, mark per-server airflow (and bypass/recirculation) as calibration parameters with priors, and report MAE before and after calibration.
• Use about 1 C MAE as a realistic twin-grade target.
• Flag uncalibrated design models as possibly off by up to about 5 C.
• Prefer vendor airflow-per-kW data in the server catalog to reduce unknowns.

Who should read it

Engineers and researchers building data-hall CFD models and digital twins and calibrating them in operation.

Related standards

• ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments

Publication details

Authors

Ruihang Wang; Xin Zhou; Linsen Dong; Yonggang Wen; Rui Tan; Li Chen; Guan Wang; Feng Zeng

Publisher

Proceedings of the 7th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (ACM)

Year

2020

Document number

arXiv:2001.10681

Pages

200-209

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.1145/3408308.3427982

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation