AI-driven Thermal-aware Data Center Capacity Planning

Overview

This document is a 2026 arXiv preprint (arXiv:2610.02442) by Cadence Design Systems engineers Yixing Li, Mark Fenton, Matthew Kaufeler, Ka Ming Leung, Xin Ai and Zhiyu Zeng. The Reality DC (formerly 6SigmaDCX) team describes a U-Net surrogate that predicts the cabinet-inlet temperature field, rather than the whole room, from multi-channel inputs: cabinet layout, IT placement and power, and ACU setpoints and fan speeds.

The model is trained on scripted batches of CFD runs, learning from CFD-extracted cabinet-inlet temperature, velocity and pressure; inlet temperature was chosen as the most operationally critical output. As a practical guideline, an initial dataset is about 500 simulation cases with around ten randomised parameters, extended as needed. The example is a 10.8 m x 9 m x 4.5 m raised-floor room with 4 rows (32 cabinets, 21 x 2U units each), four row-end ACUs and full cold-aisle containment; IT unit power was sampled from 0 to 10 kW and ACU fan speed was varied.

Results are summarised in speed and accuracy: 1000 CFD cases were used (85/10/5 split, about 5 h of training), inference took 9 ms per case against about 25 minutes for full CFD (about 150,000x faster). The paper claims MAE below 0.25 C for cabinet-inlet prediction, but a few cabinets showed maximum-inlet errors of about 3 C, while the relative temperature ranking across cabinets was preserved. The authors note that an earlier HPE 3D-CNN surrogate reported maximum absolute errors of several degrees, which motivated focusing on cabinet inlets.

The preserved ranking matters because the capacity-placement engine uses it to choose cabinets for new load. A limit is that the example covers a single room template; the results are the authors' own report and should not be generalised beyond that configuration. The method should therefore be seen as a screening stage that speeds up candidate selection, not as a replacement for full CFD.

Key points

1. The surrogate is trained on CFD-extracted cabinet-inlet temperature, velocity and pressure; inlet temperature was chosen as the most operationally critical output.
2. As a practical guideline, an initial dataset is about 500 simulation cases with around ten randomised parameters.
3. Raised-floor case: 10.8 x 9 x 4.5 m room, 4 rows (32 cabinets), four row-end ACUs, full cold-aisle containment; IT unit power 0-10 kW.
4. 1000 CFD cases (85/10/5 split, about 5 h training); inference 9 ms per case versus about 25 minutes for full CFD (about 150,000x faster).
5. MAE below 0.25 C claimed for cabinet-inlet prediction; a few cabinets showed maximum errors of about 3 C, while relative ranking was preserved.
6. An earlier HPE 3D-CNN surrogate reported maximum absolute errors of several degrees, motivating the focus on cabinet inlets.

Using it in design

• For layout iteration and add-load what-ifs, a surrogate that predicts only cabinet-inlet temperature (the compliance quantity) is cheaper and more accurate than full-field surrogates.
• Use it to rank candidate cabinets, but confirm the chosen design with full CFD because per-cabinet errors can reach about 3 C.
• Plan for about 500-1000 CFD runs per room template to train one.
• The UI should show surrogate confidence and the CFD-confirmed status.

Who should read it

Engineers and researchers building data center capacity planning, thermal simulation and CFD-based placement tools.

Related standards

• ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments

Publication details

Authors

Yixing Li; Mark Fenton; Matthew Kaufeler; Ka Ming Leung; Xin Ai; Zhiyu Zeng

Publisher

arXiv preprint (Cadence Design Systems)

Year

2026

Document number

arXiv:2610.02442

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2610.02442

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation