Fast 3D Surrogate Modeling for Data Center Thermal Management

Overview

This paper (AAAI 2025, vol. 40, no. 46, pp. 39201-39210; arXiv:2511.11722) trains 3D vision surrogates (U-Net variants, Swin UNETR and a 3D Fourier Neural Operator). The models map a voxelised data hall straight to its steady-state 3D temperature field and need no mesh or boundary-condition setup. The input is a voxel tensor with channels encoding IT power (heat sources), CRAC/vent flow and CRAC setpoint/fan speed; the grid has about 15 cm resolution (about 6.6 voxels per metre).

Training data come from close to 10,000 CFD runs generated in 6SigmaDCX across seven room configurations, each taking 11 to 35 minutes. Each cabinet was modelled as three segments of four HPE DL380 servers, with segment utilisation sampled at 25%, 60% or 90% by Latin hypercube sampling. The surrogates return a result in about 100 ms against about 35 minutes for CFD; the abstract states about a 20,000x speed-up.

Prediction errors concentrate at physical boundaries such as rack and wall edges. The surrogate was coupled to a genetic algorithm for workload placement and reduced hotspots by 7.70% compared with random workload distribution. The work is valid within its own training envelope; outside it, or for final sign-off, full CFD is needed.

Key points

1. Input voxels encode three attributes: normalised heat-source power, vent and CRAC flow rates on single-voxel slices, and CRAC setpoint/fan speed; the grid uses about 6.6 voxels per metre (about 15 cm).
2. Close to 10,000 CFD samples were generated in 6SigmaDCX across seven configurations, with runs taking 11 to 35 minutes.
3. Each cabinet was modelled as three segments of four HPE DL380 servers, with segment utilisation sampled at 25%, 60% or 90% using Latin hypercube sampling.
4. Surrogate inference took about 100 ms against about 35 minutes for CFD; the abstract states about a 20,000x speed-up.
5. Prediction errors were concentrated at physical boundaries such as rack and wall edges.
6. A genetic algorithm using the surrogate for workload placement reduced hotspots by 7.70% compared with random distribution.
7. The models are U-Net variants, Swin UNETR and a 3D Fourier Neural Operator; no mesh or boundary-condition setup is needed.

Using it in design

• Offer a fast-preview thermal mode with a voxel surrogate trained on the tool's own CFD runs: inputs of rack power by segment, CRAH setpoints, fan speeds and vent locations; output a 3D inlet-temperature map in under a second for layout iteration.
• Show the training envelope in results and fall back to full CFD for final sign-off, because errors concentrate near boundaries.
• Store parametric CFD runs from projects to build training sets.

Who should read it

Engineers and researchers seeking CFD alternatives and fast layout iteration in data centre thermal design.

Publication details

Authors

Soumyendu Sarkar; Antonio Guillen-Perez; Zachariah J Carmichael; Avisek Naug; Refik Mert Cam; Vineet Gundecha; Ashwin Ramesh Babu; Sahand Ghorbanpour; Ricardo Luna Gutierrez

Publisher

Proceedings of the AAAI Conference on Artificial Intelligence (Association for the Advancement of Artificial Intelligence (AAAI))

Year

2025

Document number

arXiv:2511.11722

Volume

40

Issue

46

Pages

39201-39210

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.1609/aaai.v40i46.41268

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation